2020-01-07 08:45:22 +00:00
|
|
|
/* GStreamer
|
|
|
|
* Copyright (C) 2020 Seungha Yang <seungha.yang@navercorp.com>
|
|
|
|
* Copyright (C) 2020 Seungha Yang <seungha@centricular.com>
|
|
|
|
*
|
|
|
|
* This library is free software; you can 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 <gst/gst.h>
|
2022-01-27 16:45:30 +00:00
|
|
|
#include "gstmfvideoencoder.h"
|
2020-08-12 11:31:32 +00:00
|
|
|
#include "gstmfvideobuffer.h"
|
2021-10-15 16:15:06 +00:00
|
|
|
#include "gstmfplatloader.h"
|
|
|
|
#include <wrl.h>
|
2020-12-26 11:39:07 +00:00
|
|
|
#include <string.h>
|
2021-06-22 16:43:08 +00:00
|
|
|
#include <cmath>
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2021-03-20 07:26:21 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
#include <d3d10.h>
|
|
|
|
#endif
|
|
|
|
|
2021-07-03 13:56:48 +00:00
|
|
|
/* *INDENT-OFF* */
|
2020-01-07 08:45:22 +00:00
|
|
|
using namespace Microsoft::WRL;
|
2022-01-23 17:41:29 +00:00
|
|
|
/* *INDENT-ON* */
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
GST_DEBUG_CATEGORY_EXTERN (gst_mf_video_encoder_debug);
|
|
|
|
#define GST_CAT_DEFAULT gst_mf_video_encoder_debug
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
#define gst_mf_video_encoder_parent_class parent_class
|
|
|
|
G_DEFINE_ABSTRACT_TYPE (GstMFVideoEncoder, gst_mf_video_encoder,
|
2020-12-26 11:39:07 +00:00
|
|
|
GST_TYPE_VIDEO_ENCODER);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
static void gst_mf_video_encoder_dispose (GObject * object);
|
|
|
|
static void gst_mf_video_encoder_set_context (GstElement * element,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstContext * context);
|
2022-01-27 16:45:30 +00:00
|
|
|
static gboolean gst_mf_video_encoder_open (GstVideoEncoder * enc);
|
|
|
|
static gboolean gst_mf_video_encoder_close (GstVideoEncoder * enc);
|
|
|
|
static gboolean gst_mf_video_encoder_start (GstVideoEncoder * enc);
|
|
|
|
static gboolean gst_mf_video_encoder_set_format (GstVideoEncoder * enc,
|
2020-01-07 08:45:22 +00:00
|
|
|
GstVideoCodecState * state);
|
2022-01-27 16:45:30 +00:00
|
|
|
static GstFlowReturn gst_mf_video_encoder_handle_frame (GstVideoEncoder * enc,
|
2020-01-07 08:45:22 +00:00
|
|
|
GstVideoCodecFrame * frame);
|
2022-01-27 16:45:30 +00:00
|
|
|
static GstFlowReturn gst_mf_video_encoder_finish (GstVideoEncoder * enc);
|
|
|
|
static gboolean gst_mf_video_encoder_flush (GstVideoEncoder * enc);
|
|
|
|
static gboolean gst_mf_video_encoder_propose_allocation (GstVideoEncoder * enc,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstQuery * query);
|
2022-01-27 16:45:30 +00:00
|
|
|
static gboolean gst_mf_video_encoder_sink_query (GstVideoEncoder * enc,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstQuery * query);
|
2022-01-27 16:45:30 +00:00
|
|
|
static gboolean gst_mf_video_encoder_src_query (GstVideoEncoder * enc,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstQuery * query);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
static HRESULT gst_mf_video_on_new_sample (GstMFTransform * object,
|
2022-01-27 16:45:30 +00:00
|
|
|
IMFSample * sample, GstMFVideoEncoder * self);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_class_init (GstMFVideoEncoderClass * klass)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2020-12-20 20:11:03 +00:00
|
|
|
GObjectClass *gobject_class = G_OBJECT_CLASS (klass);
|
|
|
|
GstElementClass *element_class = GST_ELEMENT_CLASS (klass);
|
2020-01-07 08:45:22 +00:00
|
|
|
GstVideoEncoderClass *videoenc_class = GST_VIDEO_ENCODER_CLASS (klass);
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gobject_class->dispose = gst_mf_video_encoder_dispose;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
element_class->set_context =
|
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_set_context);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
videoenc_class->open = GST_DEBUG_FUNCPTR (gst_mf_video_encoder_open);
|
|
|
|
videoenc_class->close = GST_DEBUG_FUNCPTR (gst_mf_video_encoder_close);
|
|
|
|
videoenc_class->start = GST_DEBUG_FUNCPTR (gst_mf_video_encoder_start);
|
|
|
|
videoenc_class->set_format =
|
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_set_format);
|
2020-01-07 08:45:22 +00:00
|
|
|
videoenc_class->handle_frame =
|
2022-01-27 16:45:30 +00:00
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_handle_frame);
|
|
|
|
videoenc_class->finish = GST_DEBUG_FUNCPTR (gst_mf_video_encoder_finish);
|
|
|
|
videoenc_class->flush = GST_DEBUG_FUNCPTR (gst_mf_video_encoder_flush);
|
2020-12-20 20:11:03 +00:00
|
|
|
videoenc_class->propose_allocation =
|
2022-01-27 16:45:30 +00:00
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_propose_allocation);
|
|
|
|
videoenc_class->sink_query =
|
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_sink_query);
|
|
|
|
videoenc_class->src_query =
|
|
|
|
GST_DEBUG_FUNCPTR (gst_mf_video_encoder_src_query);
|
|
|
|
|
|
|
|
gst_type_mark_as_plugin_api (GST_TYPE_MF_VIDEO_ENCODER,
|
|
|
|
(GstPluginAPIFlags) 0);
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_init (GstMFVideoEncoder * self)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
|
|
|
}
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_dispose (GObject * object)
|
2020-12-20 20:11:03 +00:00
|
|
|
{
|
|
|
|
#if GST_MF_HAVE_D3D11
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (object);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
gst_clear_object (&self->d3d11_device);
|
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
G_OBJECT_CLASS (parent_class)->dispose (object);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_set_context (GstElement * element, GstContext * context)
|
2020-12-20 20:11:03 +00:00
|
|
|
{
|
|
|
|
#if GST_MF_HAVE_D3D11
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (element);
|
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (self);
|
|
|
|
GstMFVideoEncoderDeviceCaps *device_caps = &klass->device_caps;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2021-10-08 14:07:32 +00:00
|
|
|
if (device_caps->d3d11_aware) {
|
|
|
|
gst_d3d11_handle_set_context_for_adapter_luid (element, context,
|
|
|
|
device_caps->adapter_luid, &self->other_d3d11_device);
|
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
GST_ELEMENT_CLASS (parent_class)->set_context (element, context);
|
|
|
|
}
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_open (GstVideoEncoder * enc)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (enc);
|
|
|
|
GstMFVideoEncoderDeviceCaps *device_caps = &klass->device_caps;
|
2020-01-07 08:45:22 +00:00
|
|
|
GstMFTransformEnumParams enum_params = { 0, };
|
|
|
|
MFT_REGISTER_TYPE_INFO output_type;
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
if (device_caps->d3d11_aware) {
|
|
|
|
HRESULT hr;
|
|
|
|
ID3D11Device *device_handle;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < ID3D10Multithread > multi_thread;
|
2020-12-20 20:11:03 +00:00
|
|
|
GstD3D11Device *device;
|
|
|
|
|
2021-10-08 14:07:32 +00:00
|
|
|
if (!gst_d3d11_ensure_element_data_for_adapter_luid (GST_ELEMENT (self),
|
|
|
|
device_caps->adapter_luid, &self->other_d3d11_device)) {
|
2020-12-20 20:11:03 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Other d3d11 device is unavailable");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Create our own device with D3D11_CREATE_DEVICE_VIDEO_SUPPORT */
|
2021-10-08 14:07:32 +00:00
|
|
|
self->d3d11_device =
|
|
|
|
gst_d3d11_device_new_for_adapter_luid (device_caps->adapter_luid,
|
2020-12-20 20:11:03 +00:00
|
|
|
D3D11_CREATE_DEVICE_VIDEO_SUPPORT);
|
|
|
|
if (!self->d3d11_device) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't create internal d3d11 device");
|
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
device = self->d3d11_device;
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
hr = GstMFCreateDXGIDeviceManager (&self->reset_token,
|
|
|
|
&self->device_manager);
|
2020-12-20 20:11:03 +00:00
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't create DXGI device manager");
|
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
gst_clear_object (&self->d3d11_device);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
device_handle = gst_d3d11_device_get_device_handle (device);
|
2021-03-20 07:26:21 +00:00
|
|
|
/* Enable multi thread protection as this device will be shared with
|
|
|
|
* MFT */
|
|
|
|
hr = device_handle->QueryInterface (IID_PPV_ARGS (&multi_thread));
|
|
|
|
if (!gst_d3d11_result (hr, device)) {
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"device doesn't suport ID3D10Multithread interface");
|
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
gst_clear_object (&self->d3d11_device);
|
|
|
|
}
|
|
|
|
|
|
|
|
multi_thread->SetMultithreadProtected (TRUE);
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
hr = self->device_manager->ResetDevice ((IUnknown *) device_handle,
|
|
|
|
self->reset_token);
|
|
|
|
if (!gst_mf_result (hr)) {
|
2021-07-03 13:56:48 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Couldn't reset device with given d3d11 device");
|
2020-12-20 20:11:03 +00:00
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
gst_clear_object (&self->d3d11_device);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
output_type.guidMajorType = MFMediaType_Video;
|
|
|
|
output_type.guidSubtype = klass->codec_id;
|
|
|
|
|
|
|
|
enum_params.category = MFT_CATEGORY_VIDEO_ENCODER;
|
|
|
|
enum_params.enum_flags = klass->enum_flags;
|
|
|
|
enum_params.output_typeinfo = &output_type;
|
|
|
|
enum_params.device_index = klass->device_index;
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
if (device_caps->d3d11_aware)
|
2021-10-08 14:07:32 +00:00
|
|
|
enum_params.adapter_luid = device_caps->adapter_luid;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
GST_DEBUG_OBJECT (self,
|
|
|
|
"Create MFT with enum flags: 0x%x, device index: %d, d3d11 aware: %d, "
|
|
|
|
"adapter-luid %" G_GINT64_FORMAT, klass->enum_flags, klass->device_index,
|
2021-10-08 14:07:32 +00:00
|
|
|
device_caps->d3d11_aware, device_caps->adapter_luid);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
self->transform = gst_mf_transform_new (&enum_params);
|
2022-01-27 16:45:30 +00:00
|
|
|
if (!self->transform) {
|
2020-01-07 08:45:22 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Cannot create MFT object");
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
|
|
|
}
|
2020-01-07 08:45:22 +00:00
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
/* In case of hardware MFT, it will be running on async mode.
|
|
|
|
* And new output sample callback will be called from Media Foundation's
|
|
|
|
* internal worker queue thread */
|
|
|
|
if (self->transform &&
|
|
|
|
(enum_params.enum_flags & MFT_ENUM_FLAG_HARDWARE) ==
|
2021-07-03 13:56:48 +00:00
|
|
|
MFT_ENUM_FLAG_HARDWARE) {
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
self->async_mft = TRUE;
|
|
|
|
gst_mf_transform_set_new_sample_callback (self->transform,
|
2021-07-03 13:56:48 +00:00
|
|
|
(GstMFTransformNewSampleCallback) gst_mf_video_on_new_sample, self);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
} else {
|
|
|
|
self->async_mft = FALSE;
|
|
|
|
}
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
return TRUE;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_close (GstVideoEncoder * enc)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
gst_clear_object (&self->transform);
|
|
|
|
|
|
|
|
if (self->input_state) {
|
|
|
|
gst_video_codec_state_unref (self->input_state);
|
2022-01-23 17:41:29 +00:00
|
|
|
self->input_state = nullptr;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
if (self->device_manager) {
|
|
|
|
self->device_manager->Release ();
|
|
|
|
self->device_manager = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (self->mf_allocator) {
|
|
|
|
self->mf_allocator->UninitializeSampleAllocator ();
|
|
|
|
self->mf_allocator->Release ();
|
2022-01-23 17:41:29 +00:00
|
|
|
self->mf_allocator = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
gst_clear_object (&self->other_d3d11_device);
|
|
|
|
gst_clear_object (&self->d3d11_device);
|
|
|
|
#endif
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_start (GstVideoEncoder * enc)
|
2021-06-22 16:43:08 +00:00
|
|
|
{
|
|
|
|
/* Media Foundation Transform will shift PTS in case that B-frame is enabled.
|
|
|
|
* We need to adjust DTS correspondingly */
|
|
|
|
gst_video_encoder_set_min_pts (enc, GST_SECOND * 60 * 60 * 1000);
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
static gboolean
|
2022-03-26 14:48:46 +00:00
|
|
|
gst_mf_video_encoder_init_mft (GstMFVideoEncoder * self)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-03-26 14:48:46 +00:00
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (self);
|
|
|
|
GstVideoInfo *info = &self->input_state->info;
|
|
|
|
GstCaps *caps = self->input_state->caps;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFMediaType > in_type;
|
|
|
|
ComPtr < IMFMediaType > out_type;
|
2022-01-23 17:41:29 +00:00
|
|
|
GList *input_types = nullptr;
|
2020-01-07 08:45:22 +00:00
|
|
|
GList *iter;
|
|
|
|
HRESULT hr;
|
|
|
|
gint fps_n, fps_d;
|
|
|
|
|
|
|
|
GST_DEBUG_OBJECT (self, "Set format");
|
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
gst_mf_video_encoder_finish (GST_VIDEO_ENCODER (self));
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
self->mf_pts_offset = 0;
|
|
|
|
self->has_reorder_frame = FALSE;
|
|
|
|
self->last_ret = GST_FLOW_OK;
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!gst_mf_transform_open (self->transform)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Failed to open MFT");
|
|
|
|
return FALSE;
|
|
|
|
}
|
2021-04-22 06:50:15 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
2020-12-20 20:11:03 +00:00
|
|
|
if (self->device_manager) {
|
|
|
|
if (!gst_mf_transform_set_device_manager (self->transform,
|
2021-07-03 13:56:48 +00:00
|
|
|
self->device_manager)) {
|
2020-12-20 20:11:03 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set device manager");
|
|
|
|
return FALSE;
|
|
|
|
} else {
|
|
|
|
GST_DEBUG_OBJECT (self, "set device manager done");
|
|
|
|
}
|
|
|
|
}
|
2021-04-22 06:50:15 +00:00
|
|
|
#endif
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-01-26 17:26:02 +00:00
|
|
|
hr = MFCreateMediaType (&out_type);
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return FALSE;
|
|
|
|
|
|
|
|
hr = out_type->SetGUID (MF_MT_MAJOR_TYPE, MFMediaType_Video);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return FALSE;
|
|
|
|
|
|
|
|
if (klass->set_option) {
|
2020-12-26 18:16:28 +00:00
|
|
|
if (!klass->set_option (self, self->input_state, out_type.Get ())) {
|
2020-01-07 08:45:22 +00:00
|
|
|
GST_ERROR_OBJECT (self, "subclass failed to set option");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
fps_n = GST_VIDEO_INFO_FPS_N (info);
|
|
|
|
fps_d = GST_VIDEO_INFO_FPS_D (info);
|
2021-03-24 18:24:11 +00:00
|
|
|
if (fps_n <= 0 || fps_d <= 0) {
|
|
|
|
/* XXX: not sure why. NVIDIA MFT accepts 0/1 framerate, but Intel or
|
|
|
|
* Microsoft's software MFT doesn't accept 0/1 framerate.
|
|
|
|
* Need to set something meaningful value here therefore */
|
|
|
|
fps_n = 25;
|
2020-01-07 08:45:22 +00:00
|
|
|
fps_d = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFSetAttributeRatio (out_type.Get (), MF_MT_FRAME_RATE, fps_n, fps_d);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self,
|
|
|
|
"Couldn't set framerate %d/%d, hr: 0x%x", (guint) hr);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFSetAttributeSize (out_type.Get (), MF_MT_FRAME_SIZE,
|
|
|
|
GST_VIDEO_INFO_WIDTH (info), GST_VIDEO_INFO_HEIGHT (info));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self,
|
|
|
|
"Couldn't set resolution %dx%d, hr: 0x%x", GST_VIDEO_INFO_WIDTH (info),
|
|
|
|
GST_VIDEO_INFO_HEIGHT (info), (guint) hr);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2020-06-06 12:15:34 +00:00
|
|
|
hr = MFSetAttributeRatio (out_type.Get (), MF_MT_PIXEL_ASPECT_RATIO,
|
|
|
|
GST_VIDEO_INFO_PAR_N (info), GST_VIDEO_INFO_PAR_D (info));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set par %d/%d",
|
|
|
|
GST_VIDEO_INFO_PAR_N (info), GST_VIDEO_INFO_PAR_D (info));
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
hr = out_type->SetUINT32 (MF_MT_INTERLACE_MODE, MFVideoInterlace_Progressive);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self,
|
|
|
|
"Couldn't set interlace mode, hr: 0x%x", (guint) hr);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!gst_mf_transform_set_output_type (self->transform, out_type.Get ())) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set output type");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!gst_mf_transform_get_input_available_types (self->transform,
|
2021-07-03 13:56:48 +00:00
|
|
|
&input_types)) {
|
2020-01-07 08:45:22 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Couldn't get available input types");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (iter = input_types; iter; iter = g_list_next (iter)) {
|
|
|
|
GstVideoFormat format;
|
|
|
|
GUID subtype;
|
|
|
|
IMFMediaType *type = (IMFMediaType *) iter->data;
|
|
|
|
|
|
|
|
hr = type->GetGUID (MF_MT_SUBTYPE, &subtype);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
format = gst_mf_video_subtype_to_video_format (&subtype);
|
|
|
|
if (format != GST_VIDEO_INFO_FORMAT (info))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
in_type = type;
|
|
|
|
}
|
|
|
|
|
|
|
|
g_list_free_full (input_types, (GDestroyNotify) gst_mf_media_type_release);
|
|
|
|
|
|
|
|
if (!in_type) {
|
|
|
|
GST_ERROR_OBJECT (self,
|
2022-03-26 14:48:46 +00:00
|
|
|
"Couldn't convert input caps %" GST_PTR_FORMAT " to media type", caps);
|
2020-01-07 08:45:22 +00:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFSetAttributeSize (in_type.Get (), MF_MT_FRAME_SIZE,
|
|
|
|
GST_VIDEO_INFO_WIDTH (info), GST_VIDEO_INFO_HEIGHT (info));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set frame size %dx%d",
|
|
|
|
GST_VIDEO_INFO_WIDTH (info), GST_VIDEO_INFO_HEIGHT (info));
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = in_type->SetUINT32 (MF_MT_INTERLACE_MODE, MFVideoInterlace_Progressive);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self,
|
|
|
|
"Couldn't set interlace mode, hr: 0x%x", (guint) hr);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFSetAttributeRatio (in_type.Get (), MF_MT_PIXEL_ASPECT_RATIO,
|
|
|
|
GST_VIDEO_INFO_PAR_N (info), GST_VIDEO_INFO_PAR_D (info));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set par %d/%d",
|
|
|
|
GST_VIDEO_INFO_PAR_N (info), GST_VIDEO_INFO_PAR_D (info));
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFSetAttributeRatio (in_type.Get (), MF_MT_FRAME_RATE, fps_n, fps_d);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set framerate ratio %d/%d", fps_n, fps_d);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = in_type->SetUINT32 (MF_MT_DEFAULT_STRIDE,
|
|
|
|
GST_VIDEO_INFO_PLANE_STRIDE (info, 0));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set default stride");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!gst_mf_transform_set_input_type (self->transform, in_type.Get ())) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't set input media type");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
g_assert (klass->set_src_caps != nullptr);
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!klass->set_src_caps (self, self->input_state, out_type.Get ())) {
|
|
|
|
GST_ERROR_OBJECT (self, "subclass couldn't set src caps");
|
|
|
|
return FALSE;
|
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
if (self->mf_allocator) {
|
|
|
|
self->mf_allocator->UninitializeSampleAllocator ();
|
|
|
|
self->mf_allocator->Release ();
|
2022-01-23 17:41:29 +00:00
|
|
|
self->mf_allocator = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Check whether upstream is d3d11 element */
|
2022-03-26 14:48:46 +00:00
|
|
|
GstCapsFeatures *features;
|
|
|
|
ComPtr < IMFVideoSampleAllocatorEx > allocator;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
features = gst_caps_get_features (caps, 0);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
if (features &&
|
|
|
|
gst_caps_features_contains (features,
|
|
|
|
GST_CAPS_FEATURE_MEMORY_D3D11_MEMORY)) {
|
|
|
|
GST_DEBUG_OBJECT (self, "found D3D11 memory feature");
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
hr = GstMFCreateVideoSampleAllocatorEx (IID_PPV_ARGS (&allocator));
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"IMFVideoSampleAllocatorEx interface is unavailable");
|
|
|
|
}
|
|
|
|
|
|
|
|
if (allocator) {
|
|
|
|
do {
|
|
|
|
ComPtr < IMFAttributes > attr;
|
|
|
|
|
|
|
|
hr = MFCreateAttributes (&attr, 4);
|
2020-12-20 20:11:03 +00:00
|
|
|
if (!gst_mf_result (hr))
|
2022-03-26 14:48:46 +00:00
|
|
|
break;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
/* Only one buffer per sample
|
|
|
|
* (multiple sample is usually for multi-view things) */
|
|
|
|
hr = attr->SetUINT32 (GST_GUID_MF_SA_BUFFERS_PER_SAMPLE, 1);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
hr = attr->SetUINT32 (GST_GUID_MF_SA_D3D11_USAGE, D3D11_USAGE_DEFAULT);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* TODO: Check if we need to use keyed-mutex */
|
|
|
|
hr = attr->SetUINT32 (GST_GUID_MF_SA_D3D11_SHARED_WITHOUT_MUTEX, TRUE);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
hr = attr->SetUINT32 (GST_GUID_MF_SA_D3D11_BINDFLAGS,
|
|
|
|
D3D11_BIND_VIDEO_ENCODER);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
hr = allocator->SetDirectXManager (self->device_manager);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
hr = allocator->InitializeSampleAllocatorEx (
|
|
|
|
/* min samples, since we are running on async mode,
|
|
|
|
* at least 2 samples would be required */
|
|
|
|
2,
|
|
|
|
/* max samples, why 16 + 2? it's just magic number
|
|
|
|
* (H264 max dpb size 16 + our min sample size 2) */
|
|
|
|
16 + 2, attr.Get (), in_type.Get ()
|
|
|
|
);
|
|
|
|
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
break;
|
|
|
|
|
|
|
|
GST_DEBUG_OBJECT (self, "IMFVideoSampleAllocatorEx is initialized");
|
|
|
|
|
|
|
|
self->mf_allocator = allocator.Detach ();
|
|
|
|
} while (0);
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2022-03-26 14:48:46 +00:00
|
|
|
static gboolean
|
|
|
|
gst_mf_video_encoder_set_format (GstVideoEncoder * enc,
|
|
|
|
GstVideoCodecState * state)
|
|
|
|
{
|
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
|
|
|
GST_DEBUG_OBJECT (self, "Set format");
|
|
|
|
|
|
|
|
if (self->input_state)
|
|
|
|
gst_video_codec_state_unref (self->input_state);
|
|
|
|
self->input_state = gst_video_codec_state_ref (state);
|
|
|
|
|
|
|
|
return gst_mf_video_encoder_init_mft (self);
|
|
|
|
}
|
|
|
|
|
2020-08-12 11:31:32 +00:00
|
|
|
static void
|
|
|
|
gst_mf_video_buffer_free (GstVideoFrame * frame)
|
|
|
|
{
|
|
|
|
if (!frame)
|
|
|
|
return;
|
|
|
|
|
|
|
|
gst_video_frame_unmap (frame);
|
|
|
|
g_free (frame);
|
|
|
|
}
|
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_frame_needs_copy (GstVideoFrame * vframe)
|
2020-08-12 11:31:32 +00:00
|
|
|
{
|
|
|
|
/* Single plane data can be used without copy */
|
|
|
|
if (GST_VIDEO_FRAME_N_PLANES (vframe) == 1)
|
|
|
|
return FALSE;
|
|
|
|
|
|
|
|
switch (GST_VIDEO_FRAME_FORMAT (vframe)) {
|
|
|
|
case GST_VIDEO_FORMAT_I420:
|
|
|
|
{
|
|
|
|
guint8 *data, *other_data;
|
|
|
|
guint size;
|
|
|
|
|
|
|
|
/* Unexpected stride size, Media Foundation doesn't provide API for
|
|
|
|
* per plane stride information */
|
|
|
|
if (GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 0) !=
|
|
|
|
2 * GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 1) ||
|
|
|
|
GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 1) !=
|
|
|
|
GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 2)) {
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
size = GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 0) *
|
|
|
|
GST_VIDEO_FRAME_HEIGHT (vframe);
|
|
|
|
if (size + GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 0) !=
|
|
|
|
GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 1))
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 0);
|
|
|
|
other_data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 1);
|
|
|
|
if (data + size != other_data)
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
size = GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 1) *
|
|
|
|
GST_VIDEO_FRAME_COMP_HEIGHT (vframe, 1);
|
|
|
|
if (size + GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 1) !=
|
|
|
|
GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 2))
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 1);
|
|
|
|
other_data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 2);
|
|
|
|
if (data + size != other_data)
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
case GST_VIDEO_FORMAT_NV12:
|
|
|
|
case GST_VIDEO_FORMAT_P010_10LE:
|
|
|
|
case GST_VIDEO_FORMAT_P016_LE:
|
|
|
|
{
|
|
|
|
guint8 *data, *other_data;
|
|
|
|
guint size;
|
|
|
|
|
|
|
|
/* Unexpected stride size, Media Foundation doesn't provide API for
|
|
|
|
* per plane stride information */
|
|
|
|
if (GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 0) !=
|
|
|
|
GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 1)) {
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
size = GST_VIDEO_FRAME_PLANE_STRIDE (vframe, 0) *
|
|
|
|
GST_VIDEO_FRAME_HEIGHT (vframe);
|
|
|
|
|
|
|
|
/* Unexpected padding */
|
|
|
|
if (size + GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 0) !=
|
|
|
|
GST_VIDEO_FRAME_PLANE_OFFSET (vframe, 1))
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 0);
|
|
|
|
other_data = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 1);
|
|
|
|
if (data + size != other_data)
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
g_assert_not_reached ();
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
typedef struct
|
|
|
|
{
|
2021-06-22 16:43:08 +00:00
|
|
|
LONGLONG mf_pts;
|
2022-01-27 16:45:30 +00:00
|
|
|
} GstMFVideoEncoderFrameData;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_process_input (GstMFVideoEncoder * self,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstVideoCodecFrame * frame, IMFSample * sample)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (self);
|
2020-01-07 08:45:22 +00:00
|
|
|
HRESULT hr;
|
|
|
|
gboolean unset_force_keyframe = FALSE;
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderFrameData *frame_data = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
gboolean res;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
frame_data = g_new0 (GstMFVideoEncoderFrameData, 1);
|
2020-01-07 08:45:22 +00:00
|
|
|
frame_data->mf_pts = frame->pts / 100;
|
|
|
|
|
|
|
|
gst_video_codec_frame_set_user_data (frame,
|
|
|
|
frame_data, (GDestroyNotify) g_free);
|
|
|
|
|
|
|
|
hr = sample->SetSampleTime (frame_data->mf_pts);
|
|
|
|
if (!gst_mf_result (hr))
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2021-07-03 13:56:48 +00:00
|
|
|
hr = sample->
|
|
|
|
SetSampleDuration (GST_CLOCK_TIME_IS_VALID (frame->duration) ? frame->
|
|
|
|
duration / 100 : 0);
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!gst_mf_result (hr))
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (GST_VIDEO_CODEC_FRAME_IS_FORCE_KEYFRAME (frame)) {
|
2020-12-26 11:39:07 +00:00
|
|
|
if (klass->device_caps.force_keyframe) {
|
2020-01-07 08:45:22 +00:00
|
|
|
unset_force_keyframe =
|
|
|
|
gst_mf_transform_set_codec_api_uint32 (self->transform,
|
|
|
|
&CODECAPI_AVEncVideoForceKeyFrame, TRUE);
|
|
|
|
} else {
|
|
|
|
GST_WARNING_OBJECT (self, "encoder does not support force keyframe");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
/* Unlock temporary so that we can output frame from Media Foundation's
|
|
|
|
* worker thread.
|
|
|
|
* While we are processing input, MFT might notify
|
|
|
|
* METransformHaveOutput event from Media Foundation's internal worker queue
|
|
|
|
* thread. Then we will output encoded data from the thread synchroniously,
|
|
|
|
* not from streaming (this) thread */
|
|
|
|
if (self->async_mft)
|
|
|
|
GST_VIDEO_ENCODER_STREAM_UNLOCK (self);
|
2020-12-20 20:11:03 +00:00
|
|
|
res = gst_mf_transform_process_input (self->transform, sample);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
if (self->async_mft)
|
2021-07-03 13:56:48 +00:00
|
|
|
GST_VIDEO_ENCODER_STREAM_LOCK (self);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (unset_force_keyframe) {
|
|
|
|
gst_mf_transform_set_codec_api_uint32 (self->transform,
|
|
|
|
&CODECAPI_AVEncVideoForceKeyFrame, FALSE);
|
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
if (!res) {
|
|
|
|
GST_ERROR_OBJECT (self, "Failed to process input");
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
}
|
|
|
|
|
2020-08-12 11:31:32 +00:00
|
|
|
return TRUE;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static GstVideoCodecFrame *
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_find_output_frame (GstMFVideoEncoder * self,
|
|
|
|
LONGLONG mf_pts)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
|
|
|
GList *l, *walk = gst_video_encoder_get_frames (GST_VIDEO_ENCODER (self));
|
2022-01-23 17:41:29 +00:00
|
|
|
GstVideoCodecFrame *ret = nullptr;
|
|
|
|
GstVideoCodecFrame *closest = nullptr;
|
2021-06-22 16:43:08 +00:00
|
|
|
LONGLONG min_pts_abs_diff = 0;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
for (l = walk; l; l = l->next) {
|
|
|
|
GstVideoCodecFrame *frame = (GstVideoCodecFrame *) l->data;
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderFrameData *data = (GstMFVideoEncoderFrameData *)
|
2020-01-07 08:45:22 +00:00
|
|
|
gst_video_codec_frame_get_user_data (frame);
|
2021-06-22 16:43:08 +00:00
|
|
|
LONGLONG abs_diff;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (!data)
|
|
|
|
continue;
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
if (mf_pts == data->mf_pts) {
|
2020-01-07 08:45:22 +00:00
|
|
|
ret = frame;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
abs_diff = std::abs (mf_pts - data->mf_pts);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
if (!closest || abs_diff < min_pts_abs_diff) {
|
|
|
|
closest = frame;
|
|
|
|
min_pts_abs_diff = abs_diff;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
if (!ret && closest)
|
|
|
|
ret = closest;
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
if (ret) {
|
|
|
|
gst_video_codec_frame_ref (ret);
|
|
|
|
} else {
|
2021-06-22 16:43:08 +00:00
|
|
|
/* XXX: Shouldn't happen, but possible if no GstVideoCodecFrame holds
|
|
|
|
* user data for some reasons */
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"Failed to find closest GstVideoCodecFrame with MF pts %"
|
|
|
|
G_GINT64_FORMAT, mf_pts);
|
2020-01-07 08:45:22 +00:00
|
|
|
ret = gst_video_encoder_get_oldest_frame (GST_VIDEO_ENCODER (self));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (walk)
|
|
|
|
g_list_free_full (walk, (GDestroyNotify) gst_video_codec_frame_unref);
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
static HRESULT
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_finish_sample (GstMFVideoEncoder * self,
|
|
|
|
IMFSample * sample)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
HRESULT hr = S_OK;
|
2020-01-07 08:45:22 +00:00
|
|
|
BYTE *data;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFMediaBuffer > media_buffer;
|
2020-01-07 08:45:22 +00:00
|
|
|
GstBuffer *buffer;
|
|
|
|
GstFlowReturn res = GST_FLOW_ERROR;
|
|
|
|
GstVideoCodecFrame *frame;
|
|
|
|
LONGLONG sample_timestamp;
|
|
|
|
LONGLONG sample_duration;
|
2021-06-22 16:43:08 +00:00
|
|
|
LONGLONG target_mf_pts;
|
|
|
|
UINT64 mf_dts;
|
2020-01-07 08:45:22 +00:00
|
|
|
UINT32 keyframe = FALSE;
|
|
|
|
DWORD buffer_len;
|
2021-06-22 16:43:08 +00:00
|
|
|
GstClockTime pts, dts, duration;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-26 17:26:02 +00:00
|
|
|
hr = sample->GetBufferByIndex (0, &media_buffer);
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!gst_mf_result (hr))
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
goto done;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
hr = media_buffer->Lock (&data, nullptr, &buffer_len);
|
2020-01-07 08:45:22 +00:00
|
|
|
if (!gst_mf_result (hr))
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
goto done;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
buffer = gst_buffer_new_allocate (nullptr, buffer_len, nullptr);
|
2020-01-07 08:45:22 +00:00
|
|
|
gst_buffer_fill (buffer, 0, data, buffer_len);
|
|
|
|
media_buffer->Unlock ();
|
|
|
|
|
|
|
|
sample->GetSampleTime (&sample_timestamp);
|
2021-06-22 16:43:08 +00:00
|
|
|
target_mf_pts = sample_timestamp;
|
2020-01-07 08:45:22 +00:00
|
|
|
sample->GetSampleDuration (&sample_duration);
|
|
|
|
sample->GetUINT32 (MFSampleExtension_CleanPoint, &keyframe);
|
|
|
|
|
|
|
|
hr = sample->GetUINT64 (MFSampleExtension_DecodeTimestamp, &mf_dts);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
if (FAILED (hr)) {
|
2020-01-07 08:45:22 +00:00
|
|
|
mf_dts = sample_timestamp;
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
hr = S_OK;
|
|
|
|
}
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
pts = sample_timestamp * 100;
|
|
|
|
dts = mf_dts * 100;
|
|
|
|
duration = sample_duration * 100;
|
|
|
|
|
|
|
|
GST_LOG_OBJECT (self, "Finish sample, MF pts %" GST_TIME_FORMAT " MF dts %"
|
|
|
|
GST_TIME_FORMAT ", MF duration %" GST_TIME_FORMAT,
|
|
|
|
GST_TIME_ARGS (pts), GST_TIME_ARGS (dts), GST_TIME_ARGS (duration));
|
|
|
|
|
|
|
|
/* NOTE: When B-frame is enabled, MFT shows following pattern
|
|
|
|
* (input timestamp starts from 1000:00:00.000000000, and 30fps)
|
|
|
|
*
|
|
|
|
* Frame-1: MF pts 0:00.033333300 MF dts 0:00.000000000
|
|
|
|
* Frame-2: MF pts 0:00.133333300 MF dts 0:00.033333300
|
|
|
|
* Frame-3: MF pts 0:00.066666600 MF dts 0:00.066666600
|
|
|
|
* Frame-4: MF pts 0:00.099999900 MF dts 0:00.100000000
|
|
|
|
*
|
|
|
|
* - Sounds MFT doesn't support negative timestamp, so PTS of each frame seems
|
|
|
|
* to be shifthed
|
|
|
|
* - DTS is likely based on timestamp we've set to input sample,
|
|
|
|
* but some frames has (especially Frame-4 case) unexpected PTS and
|
|
|
|
* even PTS < DTS. That would be the result of PTS shifting
|
|
|
|
*
|
|
|
|
* To handle this case,
|
|
|
|
* - Calculate timestamp offset "Frame-1 PTS" - "Frame-1 DTS" (== duration),
|
|
|
|
* and compensate PTS/DTS of each frame
|
|
|
|
* - Needs additional offset for DTS to compenstate GST/MF timescale difference
|
|
|
|
* (MF uses 100ns timescale). So DTS offset should be "PTS offset + 100ns"
|
|
|
|
* - Find corresponding GstVideoCodecFrame by using compensated PTS.
|
|
|
|
* Note that MFT doesn't support user-data for tracing input/output sample
|
|
|
|
* pair. So, timestamp based lookup is the only way to map MF sample
|
|
|
|
* and our GstVideoCodecFrame
|
|
|
|
*/
|
|
|
|
if (self->has_reorder_frame) {
|
|
|
|
/* This would be the first frame */
|
|
|
|
if (self->mf_pts_offset == 0) {
|
|
|
|
LONGLONG mf_pts_offset = -1;
|
|
|
|
if (sample_timestamp > mf_dts) {
|
|
|
|
mf_pts_offset = sample_timestamp - mf_dts;
|
|
|
|
GST_DEBUG_OBJECT (self, "Calculates PTS offset using \"PTS - DTS\": %"
|
|
|
|
G_GINT64_FORMAT, mf_pts_offset);
|
|
|
|
} else if (sample_duration > 0) {
|
|
|
|
mf_pts_offset = sample_duration;
|
|
|
|
GST_DEBUG_OBJECT (self, "Calculates PTS offset using duration: %"
|
|
|
|
G_GINT64_FORMAT, mf_pts_offset);
|
|
|
|
} else {
|
|
|
|
GST_WARNING_OBJECT (self, "Cannot calculate PTS offset");
|
|
|
|
}
|
|
|
|
|
|
|
|
self->mf_pts_offset = mf_pts_offset;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (self->mf_pts_offset > 0) {
|
|
|
|
target_mf_pts -= self->mf_pts_offset;
|
|
|
|
|
|
|
|
pts -= (self->mf_pts_offset * 100);
|
|
|
|
/* +1 to compensate timescale difference */
|
|
|
|
dts -= ((self->mf_pts_offset + 1) * 100);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
frame = gst_mf_video_encoder_find_output_frame (self, target_mf_pts);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (frame) {
|
|
|
|
if (keyframe) {
|
|
|
|
GST_DEBUG_OBJECT (self, "Keyframe pts %" GST_TIME_FORMAT,
|
|
|
|
GST_TIME_ARGS (frame->pts));
|
|
|
|
GST_VIDEO_CODEC_FRAME_SET_SYNC_POINT (frame);
|
|
|
|
}
|
|
|
|
|
|
|
|
frame->output_buffer = buffer;
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
/* Update DTS only if B-frame was enabled, but use input frame pts as-is.
|
|
|
|
* Otherwise we will lost at most 100ns precision */
|
|
|
|
if (self->has_reorder_frame) {
|
|
|
|
frame->dts = dts;
|
|
|
|
} else {
|
|
|
|
frame->dts = frame->pts;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* make sure PTS > DTS */
|
|
|
|
if (GST_CLOCK_TIME_IS_VALID (frame->pts) &&
|
2021-07-03 13:56:48 +00:00
|
|
|
GST_CLOCK_TIME_IS_VALID (frame->dts) && frame->pts < frame->dts) {
|
2021-06-22 16:43:08 +00:00
|
|
|
GST_WARNING_OBJECT (self, "Calculated DTS %" GST_TIME_FORMAT
|
|
|
|
" is larger than PTS %" GST_TIME_FORMAT, GST_TIME_ARGS (frame->pts),
|
|
|
|
GST_TIME_ARGS (frame->dts));
|
|
|
|
|
|
|
|
/* XXX: just set clock-time-none? */
|
|
|
|
frame->dts = frame->pts;
|
|
|
|
}
|
|
|
|
|
|
|
|
GST_LOG_OBJECT (self, "Frame pts %" GST_TIME_FORMAT ", Frame DTS %"
|
2021-07-03 13:56:48 +00:00
|
|
|
GST_TIME_FORMAT, GST_TIME_ARGS (frame->pts),
|
|
|
|
GST_TIME_ARGS (frame->dts));
|
2021-06-22 16:43:08 +00:00
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
res = gst_video_encoder_finish_frame (GST_VIDEO_ENCODER (self), frame);
|
|
|
|
} else {
|
2021-06-22 16:43:08 +00:00
|
|
|
GST_BUFFER_PTS (buffer) = pts;
|
|
|
|
GST_BUFFER_DTS (buffer) = dts;
|
|
|
|
GST_BUFFER_DURATION (buffer) = duration;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (keyframe) {
|
|
|
|
GST_DEBUG_OBJECT (self, "Keyframe pts %" GST_TIME_FORMAT,
|
|
|
|
GST_BUFFER_PTS (buffer));
|
|
|
|
GST_BUFFER_FLAG_UNSET (buffer, GST_BUFFER_FLAG_DELTA_UNIT);
|
|
|
|
} else {
|
|
|
|
GST_BUFFER_FLAG_SET (buffer, GST_BUFFER_FLAG_DELTA_UNIT);
|
|
|
|
}
|
|
|
|
|
2021-06-22 16:43:08 +00:00
|
|
|
GST_LOG_OBJECT (self, "Buffer pts %" GST_TIME_FORMAT ", Buffer DTS %"
|
|
|
|
GST_TIME_FORMAT, GST_TIME_ARGS (pts), GST_TIME_ARGS (dts));
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
res = gst_pad_push (GST_VIDEO_ENCODER_SRC_PAD (self), buffer);
|
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
done:
|
|
|
|
self->last_ret = res;
|
|
|
|
|
|
|
|
return hr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static GstFlowReturn
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_process_output (GstMFVideoEncoder * self)
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
{
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFSample > sample;
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
GstFlowReturn res = GST_FLOW_ERROR;
|
|
|
|
|
|
|
|
res = gst_mf_transform_get_output (self->transform, &sample);
|
|
|
|
|
|
|
|
if (res != GST_FLOW_OK)
|
|
|
|
return res;
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_finish_sample (self, sample.Get ());
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
|
|
|
|
return self->last_ret;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_create_input_sample (GstMFVideoEncoder * self,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstVideoCodecFrame * frame, IMFSample ** sample)
|
|
|
|
{
|
|
|
|
HRESULT hr;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFSample > new_sample;
|
|
|
|
ComPtr < IMFMediaBuffer > media_buffer;
|
|
|
|
ComPtr < IGstMFVideoBuffer > video_buffer;
|
2020-12-20 20:11:03 +00:00
|
|
|
GstVideoInfo *info = &self->input_state->info;
|
|
|
|
gint i, j;
|
2022-01-23 17:41:29 +00:00
|
|
|
GstVideoFrame *vframe = nullptr;
|
|
|
|
BYTE *data = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
gboolean need_copy;
|
|
|
|
|
|
|
|
vframe = g_new0 (GstVideoFrame, 1);
|
|
|
|
|
|
|
|
if (!gst_video_frame_map (vframe, info, frame->input_buffer, GST_MAP_READ)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't map input frame");
|
|
|
|
g_free (vframe);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = MFCreateSample (&new_sample);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
|
|
|
|
/* Check if we can forward this memory to Media Foundation without copy */
|
2022-01-27 16:45:30 +00:00
|
|
|
need_copy = gst_mf_video_encoder_frame_needs_copy (vframe);
|
2020-12-20 20:11:03 +00:00
|
|
|
if (need_copy) {
|
|
|
|
GST_TRACE_OBJECT (self, "Copy input buffer into Media Foundation memory");
|
|
|
|
hr = MFCreateMemoryBuffer (GST_VIDEO_INFO_SIZE (info), &media_buffer);
|
|
|
|
} else {
|
|
|
|
GST_TRACE_OBJECT (self, "Can use input buffer without copy");
|
|
|
|
hr = IGstMFVideoBuffer::CreateInstanceWrapped (&vframe->info,
|
2021-07-03 13:56:48 +00:00
|
|
|
(BYTE *) GST_VIDEO_FRAME_PLANE_DATA (vframe, 0),
|
|
|
|
GST_VIDEO_INFO_SIZE (&vframe->info), &media_buffer);
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
|
|
|
|
if (!need_copy) {
|
|
|
|
hr = media_buffer.As (&video_buffer);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
} else {
|
2022-01-23 17:41:29 +00:00
|
|
|
hr = media_buffer->Lock (&data, nullptr, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
|
|
|
|
for (i = 0; i < GST_VIDEO_INFO_N_PLANES (info); i++) {
|
|
|
|
guint8 *src, *dst;
|
|
|
|
gint src_stride, dst_stride;
|
|
|
|
gint width;
|
|
|
|
|
|
|
|
src = (guint8 *) GST_VIDEO_FRAME_PLANE_DATA (vframe, i);
|
|
|
|
dst = data + GST_VIDEO_INFO_PLANE_OFFSET (info, i);
|
|
|
|
|
|
|
|
src_stride = GST_VIDEO_FRAME_PLANE_STRIDE (vframe, i);
|
|
|
|
dst_stride = GST_VIDEO_INFO_PLANE_STRIDE (info, i);
|
|
|
|
|
|
|
|
width = GST_VIDEO_INFO_COMP_WIDTH (info, i)
|
|
|
|
* GST_VIDEO_INFO_COMP_PSTRIDE (info, i);
|
|
|
|
|
|
|
|
for (j = 0; j < GST_VIDEO_INFO_COMP_HEIGHT (info, i); j++) {
|
|
|
|
memcpy (dst, src, width);
|
|
|
|
src += src_stride;
|
|
|
|
dst += dst_stride;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
media_buffer->Unlock ();
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = media_buffer->SetCurrentLength (GST_VIDEO_INFO_SIZE (info));
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
|
|
|
|
hr = new_sample->AddBuffer (media_buffer.Get ());
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
|
|
|
|
if (!need_copy) {
|
|
|
|
/* IGstMFVideoBuffer will hold GstVideoFrame (+ GstBuffer), then it will be
|
|
|
|
* cleared when it's no more referenced by Media Foundation internals */
|
|
|
|
hr = video_buffer->SetUserData ((gpointer) vframe,
|
|
|
|
(GDestroyNotify) gst_mf_video_buffer_free);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
goto error;
|
|
|
|
} else {
|
|
|
|
gst_video_frame_unmap (vframe);
|
|
|
|
g_free (vframe);
|
2022-01-23 17:41:29 +00:00
|
|
|
vframe = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
*sample = new_sample.Detach ();
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
error:
|
|
|
|
if (vframe) {
|
|
|
|
gst_video_frame_unmap (vframe);
|
|
|
|
g_free (vframe);
|
|
|
|
}
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_create_input_sample_d3d11 (GstMFVideoEncoder * self,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstVideoCodecFrame * frame, IMFSample ** sample)
|
|
|
|
{
|
2022-03-26 12:47:10 +00:00
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (self);
|
|
|
|
GstMFVideoEncoderDeviceCaps *device_caps = &klass->device_caps;
|
2020-12-20 20:11:03 +00:00
|
|
|
HRESULT hr;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFSample > new_sample;
|
|
|
|
ComPtr < IMFMediaBuffer > mf_buffer;
|
|
|
|
ComPtr < IMFDXGIBuffer > dxgi_buffer;
|
|
|
|
ComPtr < ID3D11Texture2D > mf_texture;
|
|
|
|
ComPtr < IDXGIResource > dxgi_resource;
|
|
|
|
ComPtr < ID3D11Texture2D > shared_texture;
|
|
|
|
ComPtr < ID3D11Query > query;
|
2020-12-20 20:11:03 +00:00
|
|
|
D3D11_QUERY_DESC query_desc;
|
|
|
|
BOOL sync_done = FALSE;
|
|
|
|
HANDLE shared_handle;
|
|
|
|
GstMemory *mem;
|
|
|
|
GstD3D11Memory *dmem;
|
|
|
|
ID3D11Texture2D *texture;
|
|
|
|
ID3D11Device *device_handle;
|
|
|
|
ID3D11DeviceContext *context_handle;
|
|
|
|
GstMapInfo info;
|
|
|
|
D3D11_BOX src_box = { 0, };
|
|
|
|
D3D11_TEXTURE2D_DESC dst_desc, src_desc;
|
|
|
|
guint subidx;
|
2022-03-26 12:47:10 +00:00
|
|
|
gint64 adapter_luid;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
if (!self->mf_allocator) {
|
|
|
|
GST_WARNING_OBJECT (self, "IMFVideoSampleAllocatorEx was configured");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
mem = gst_buffer_peek_memory (frame->input_buffer, 0);
|
|
|
|
if (!gst_is_d3d11_memory (mem)) {
|
|
|
|
GST_WARNING_OBJECT (self, "Non-d3d11 memory");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2022-03-26 12:47:10 +00:00
|
|
|
dmem = GST_D3D11_MEMORY_CAST (mem);
|
|
|
|
g_object_get (dmem->device, "adapter-luid", &adapter_luid, nullptr);
|
|
|
|
if (adapter_luid != device_caps->adapter_luid) {
|
|
|
|
GST_LOG_OBJECT (self, "Buffer from different GPU");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
device_handle = gst_d3d11_device_get_device_handle (dmem->device);
|
|
|
|
context_handle = gst_d3d11_device_get_device_context_handle (dmem->device);
|
|
|
|
|
|
|
|
/* 1) Allocate new encoding surface */
|
|
|
|
hr = self->mf_allocator->AllocateSample (&new_sample);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"Couldn't allocate new sample via IMFVideoSampleAllocatorEx");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = new_sample->GetBufferByIndex (0, &mf_buffer);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self, "Couldn't get IMFMediaBuffer from sample");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = mf_buffer.As (&dxgi_buffer);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self, "Couldn't get IMFDXGIBuffer from IMFMediaBuffer");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = dxgi_buffer->GetResource (IID_PPV_ARGS (&mf_texture));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"Couldn't get ID3D11Texture2D from IMFDXGIBuffer");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = mf_texture.As (&dxgi_resource);
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self,
|
|
|
|
"Couldn't get IDXGIResource from ID3D11Texture2D");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
hr = dxgi_resource->GetSharedHandle (&shared_handle);
|
|
|
|
if (!gst_mf_result (hr)) {
|
2021-07-03 13:56:48 +00:00
|
|
|
GST_WARNING_OBJECT (self, "Couldn't get shared handle from IDXGIResource");
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Allocation succeeded. Now open shared texture to access it from
|
|
|
|
* other device */
|
|
|
|
hr = device_handle->OpenSharedResource (shared_handle,
|
|
|
|
IID_PPV_ARGS (&shared_texture));
|
|
|
|
if (!gst_mf_result (hr)) {
|
|
|
|
GST_WARNING_OBJECT (self, "Couldn't open shared resource");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 2) Copy upstream texture to mf's texture */
|
|
|
|
/* Map memory so that ensure pending upload from staging texture */
|
2021-07-03 13:56:48 +00:00
|
|
|
if (!gst_memory_map (mem, &info,
|
|
|
|
(GstMapFlags) (GST_MAP_READ | GST_MAP_D3D11))) {
|
2020-12-20 20:11:03 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Couldn't map d3d11 memory");
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
texture = (ID3D11Texture2D *) info.data;
|
|
|
|
texture->GetDesc (&src_desc);
|
|
|
|
shared_texture->GetDesc (&dst_desc);
|
|
|
|
subidx = gst_d3d11_memory_get_subresource_index (dmem);
|
|
|
|
|
|
|
|
/* src/dst texture size might be different if padding was used.
|
|
|
|
* select smaller size */
|
|
|
|
src_box.left = 0;
|
|
|
|
src_box.top = 0;
|
|
|
|
src_box.front = 0;
|
|
|
|
src_box.back = 1;
|
|
|
|
src_box.right = MIN (src_desc.Width, dst_desc.Width);
|
|
|
|
src_box.bottom = MIN (src_desc.Height, dst_desc.Height);
|
|
|
|
|
|
|
|
/* CopySubresourceRegion() might not be able to guarantee
|
|
|
|
* copying. To ensure it, we will make use of d3d11 query */
|
|
|
|
query_desc.Query = D3D11_QUERY_EVENT;
|
|
|
|
query_desc.MiscFlags = 0;
|
|
|
|
|
|
|
|
hr = device_handle->CreateQuery (&query_desc, &query);
|
|
|
|
if (!gst_d3d11_result (hr, dmem->device)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't Create event query");
|
2022-03-26 12:47:10 +00:00
|
|
|
gst_memory_unmap (mem, &info);
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2021-03-20 07:15:35 +00:00
|
|
|
gst_d3d11_device_lock (dmem->device);
|
2020-12-20 20:11:03 +00:00
|
|
|
context_handle->CopySubresourceRegion (shared_texture.Get (), 0, 0, 0, 0,
|
|
|
|
texture, subidx, &src_box);
|
2021-07-03 13:56:48 +00:00
|
|
|
context_handle->End (query.Get ());
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
/* Wait until all issued GPU commands are finished */
|
|
|
|
do {
|
2022-03-26 12:43:50 +00:00
|
|
|
hr = context_handle->GetData (query.Get (), &sync_done, sizeof (BOOL), 0);
|
2020-12-20 20:11:03 +00:00
|
|
|
} while (!sync_done && (hr == S_OK || hr == S_FALSE));
|
|
|
|
|
|
|
|
if (!gst_d3d11_result (hr, dmem->device)) {
|
|
|
|
GST_ERROR_OBJECT (self, "Couldn't sync GPU operation");
|
2021-03-20 07:15:35 +00:00
|
|
|
gst_d3d11_device_unlock (dmem->device);
|
2020-12-20 20:11:03 +00:00
|
|
|
gst_memory_unmap (mem, &info);
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2021-03-20 07:15:35 +00:00
|
|
|
gst_d3d11_device_unlock (dmem->device);
|
2020-12-20 20:11:03 +00:00
|
|
|
gst_memory_unmap (mem, &info);
|
|
|
|
|
|
|
|
*sample = new_sample.Detach ();
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
static GstFlowReturn
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_handle_frame (GstVideoEncoder * enc,
|
2020-01-07 08:45:22 +00:00
|
|
|
GstVideoCodecFrame * frame)
|
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-05-24 10:12:28 +00:00
|
|
|
GstFlowReturn ret = GST_FLOW_OK;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFSample > sample;
|
2022-03-26 14:48:46 +00:00
|
|
|
GstMFVideoEncoderClass *klass = GST_MF_VIDEO_ENCODER_GET_CLASS (self);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2021-06-30 09:11:46 +00:00
|
|
|
if (self->last_ret != GST_FLOW_OK) {
|
|
|
|
GST_DEBUG_OBJECT (self, "Last return was %s", gst_flow_get_name (ret));
|
|
|
|
ret = self->last_ret;
|
|
|
|
goto done;
|
|
|
|
}
|
2022-03-26 14:48:46 +00:00
|
|
|
|
|
|
|
if (klass->check_reconfigure (self) && !gst_mf_video_encoder_init_mft (self)) {
|
|
|
|
GST_ELEMENT_ERROR (self, STREAM, ENCODE, (nullptr),
|
|
|
|
("Failed to reconfigure encoder"));
|
|
|
|
return GST_FLOW_ERROR;
|
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
if (self->mf_allocator &&
|
2022-01-27 16:45:30 +00:00
|
|
|
!gst_mf_video_encoder_create_input_sample_d3d11 (self, frame, &sample)) {
|
2022-03-26 12:47:10 +00:00
|
|
|
GST_LOG_OBJECT (self, "Failed to create IMFSample for D3D11");
|
2020-12-20 20:11:03 +00:00
|
|
|
sample = nullptr;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
if (!sample
|
|
|
|
&& !gst_mf_video_encoder_create_input_sample (self, frame, &sample)) {
|
2020-12-20 20:11:03 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Failed to create IMFSample");
|
2021-06-30 09:11:46 +00:00
|
|
|
ret = GST_FLOW_ERROR;
|
|
|
|
goto done;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
2020-01-07 08:45:22 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
if (!gst_mf_video_encoder_process_input (self, frame, sample.Get ())) {
|
2020-01-07 08:45:22 +00:00
|
|
|
GST_ERROR_OBJECT (self, "Failed to process input");
|
2020-05-24 10:12:28 +00:00
|
|
|
ret = GST_FLOW_ERROR;
|
|
|
|
goto done;
|
2020-01-07 08:45:22 +00:00
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
/* Don't call process_output for async (hardware) MFT. We will output
|
|
|
|
* encoded data from gst_mf_video_on_new_sample() callback which is called
|
|
|
|
* from Media Foundation's internal worker queue thread */
|
|
|
|
if (!self->async_mft) {
|
|
|
|
do {
|
2022-01-27 16:45:30 +00:00
|
|
|
ret = gst_mf_video_encoder_process_output (self);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
} while (ret == GST_FLOW_OK);
|
|
|
|
}
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (ret == GST_MF_TRANSFORM_FLOW_NEED_DATA)
|
|
|
|
ret = GST_FLOW_OK;
|
|
|
|
|
2020-05-24 10:12:28 +00:00
|
|
|
done:
|
|
|
|
gst_video_codec_frame_unref (frame);
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static GstFlowReturn
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_finish (GstVideoEncoder * enc)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-01-07 08:45:22 +00:00
|
|
|
GstFlowReturn ret = GST_FLOW_OK;
|
|
|
|
|
|
|
|
if (!self->transform)
|
|
|
|
return GST_FLOW_OK;
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
/* Unlock temporary so that we can output frame from Media Foundation's
|
|
|
|
* worker thread */
|
|
|
|
if (self->async_mft)
|
|
|
|
GST_VIDEO_ENCODER_STREAM_UNLOCK (enc);
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
gst_mf_transform_drain (self->transform);
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
if (self->async_mft)
|
|
|
|
GST_VIDEO_ENCODER_STREAM_LOCK (enc);
|
|
|
|
|
|
|
|
if (!self->async_mft) {
|
|
|
|
do {
|
2022-01-27 16:45:30 +00:00
|
|
|
ret = gst_mf_video_encoder_process_output (self);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
} while (ret == GST_FLOW_OK);
|
|
|
|
}
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (ret == GST_MF_TRANSFORM_FLOW_NEED_DATA)
|
|
|
|
ret = GST_FLOW_OK;
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_flush (GstVideoEncoder * enc)
|
2020-01-07 08:45:22 +00:00
|
|
|
{
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-01-07 08:45:22 +00:00
|
|
|
|
|
|
|
if (!self->transform)
|
2021-06-30 09:11:46 +00:00
|
|
|
goto out;
|
2020-01-07 08:45:22 +00:00
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
/* Unlock while flushing, while flushing, new sample callback might happen */
|
|
|
|
if (self->async_mft)
|
|
|
|
GST_VIDEO_ENCODER_STREAM_UNLOCK (enc);
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
gst_mf_transform_flush (self->transform);
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
if (self->async_mft)
|
|
|
|
GST_VIDEO_ENCODER_STREAM_LOCK (enc);
|
|
|
|
|
2021-06-30 09:11:46 +00:00
|
|
|
out:
|
|
|
|
self->last_ret = GST_FLOW_OK;
|
|
|
|
|
2020-01-07 08:45:22 +00:00
|
|
|
return TRUE;
|
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_propose_allocation (GstVideoEncoder * enc,
|
|
|
|
GstQuery * query)
|
2020-12-20 20:11:03 +00:00
|
|
|
{
|
|
|
|
#if GST_MF_HAVE_D3D11
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-12-20 20:11:03 +00:00
|
|
|
GstVideoInfo info;
|
2022-01-23 17:41:29 +00:00
|
|
|
GstBufferPool *pool = nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
GstCaps *caps;
|
|
|
|
guint size;
|
|
|
|
GstD3D11Device *device = self->other_d3d11_device;
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
gst_query_parse_allocation (query, &caps, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
if (caps == nullptr)
|
2020-12-20 20:11:03 +00:00
|
|
|
return FALSE;
|
|
|
|
|
|
|
|
if (!gst_video_info_from_caps (&info, caps))
|
|
|
|
return FALSE;
|
|
|
|
|
|
|
|
if (gst_query_get_n_allocation_pools (query) == 0) {
|
|
|
|
GstCapsFeatures *features;
|
|
|
|
GstStructure *config;
|
|
|
|
gboolean is_d3d11 = FALSE;
|
|
|
|
|
|
|
|
features = gst_caps_get_features (caps, 0);
|
|
|
|
|
|
|
|
if (features && gst_caps_features_contains (features,
|
|
|
|
GST_CAPS_FEATURE_MEMORY_D3D11_MEMORY)) {
|
|
|
|
GST_DEBUG_OBJECT (self, "Allocation caps supports d3d11 memory");
|
|
|
|
pool = gst_d3d11_buffer_pool_new (device);
|
|
|
|
is_d3d11 = TRUE;
|
|
|
|
} else {
|
|
|
|
pool = gst_video_buffer_pool_new ();
|
|
|
|
}
|
|
|
|
|
|
|
|
config = gst_buffer_pool_get_config (pool);
|
|
|
|
|
|
|
|
gst_buffer_pool_config_add_option (config,
|
|
|
|
GST_BUFFER_POOL_OPTION_VIDEO_META);
|
|
|
|
|
|
|
|
/* d3d11 pool does not support video alignment */
|
|
|
|
if (!is_d3d11) {
|
|
|
|
gst_buffer_pool_config_add_option (config,
|
|
|
|
GST_BUFFER_POOL_OPTION_VIDEO_ALIGNMENT);
|
|
|
|
}
|
|
|
|
|
|
|
|
size = GST_VIDEO_INFO_SIZE (&info);
|
|
|
|
gst_buffer_pool_config_set_params (config, caps, size, 0, 0);
|
|
|
|
|
|
|
|
if (!gst_buffer_pool_set_config (pool, config))
|
|
|
|
goto config_failed;
|
|
|
|
|
2021-08-21 15:33:58 +00:00
|
|
|
/* d3d11 buffer pool will update buffer size based on allocated texture,
|
|
|
|
* get size from config again */
|
|
|
|
if (is_d3d11) {
|
|
|
|
config = gst_buffer_pool_get_config (pool);
|
|
|
|
gst_buffer_pool_config_get_params (config,
|
|
|
|
nullptr, &size, nullptr, nullptr);
|
|
|
|
gst_structure_free (config);
|
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
gst_query_add_allocation_pool (query, pool, size, 0, 0);
|
|
|
|
gst_object_unref (pool);
|
|
|
|
}
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
gst_query_add_allocation_meta (query, GST_VIDEO_META_API_TYPE, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
/* ERRORS */
|
|
|
|
config_failed:
|
|
|
|
{
|
|
|
|
GST_ERROR_OBJECT (self, "failed to set config");
|
|
|
|
gst_object_unref (pool);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
return GST_VIDEO_ENCODER_CLASS (parent_class)->propose_allocation (enc,
|
2021-07-03 13:56:48 +00:00
|
|
|
query);
|
2020-12-20 20:11:03 +00:00
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_sink_query (GstVideoEncoder * enc, GstQuery * query)
|
2020-12-20 20:11:03 +00:00
|
|
|
{
|
|
|
|
#if GST_MF_HAVE_D3D11
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
switch (GST_QUERY_TYPE (query)) {
|
|
|
|
case GST_QUERY_CONTEXT:
|
|
|
|
if (gst_d3d11_handle_context_query (GST_ELEMENT (self),
|
|
|
|
query, self->other_d3d11_device)) {
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
return GST_VIDEO_ENCODER_CLASS (parent_class)->sink_query (enc, query);
|
|
|
|
}
|
|
|
|
|
|
|
|
static gboolean
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_src_query (GstVideoEncoder * enc, GstQuery * query)
|
2020-12-20 20:11:03 +00:00
|
|
|
{
|
|
|
|
#if GST_MF_HAVE_D3D11
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoder *self = GST_MF_VIDEO_ENCODER (enc);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
switch (GST_QUERY_TYPE (query)) {
|
|
|
|
case GST_QUERY_CONTEXT:
|
|
|
|
if (gst_d3d11_handle_context_query (GST_ELEMENT (self),
|
|
|
|
query, self->other_d3d11_device)) {
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
return GST_VIDEO_ENCODER_CLASS (parent_class)->src_query (enc, query);
|
|
|
|
}
|
|
|
|
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
static HRESULT
|
|
|
|
gst_mf_video_on_new_sample (GstMFTransform * object,
|
2022-01-27 16:45:30 +00:00
|
|
|
IMFSample * sample, GstMFVideoEncoder * self)
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
{
|
|
|
|
GST_LOG_OBJECT (self, "New Sample callback");
|
|
|
|
|
|
|
|
/* NOTE: this callback will be called from Media Foundation's internal
|
|
|
|
* worker queue thread */
|
|
|
|
GST_VIDEO_ENCODER_STREAM_LOCK (self);
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_finish_sample (self, sample);
|
mfvideoenc: Improve latency performance for hardware encoder
Unlike software MFT (Media Foundation Transform) which is synchronous
in terms of processing input and output data, hardware MFT works
in asynchronous mode. output data might not be available right after
we pushed one input data into MFT.
Note that async MFT will fire two events, one is "METransformNeedInput"
which happens when MFT can accept more input data,
and the other is "METransformHaveOutput", that's for signaling
there's pending data which can be outputted immediately.
To listen the events, we can wait synchronously via
IMFMediaEventGenerator::GetEvent() or make use of IMFAsyncCallback
object which is asynchronous way and the event will be notified
from Media Foundation's internal worker queue thread.
To handle such asynchronous operation, previous working flow was
as follows (IMFMediaEventGenerator::GetEvent() was used for now)
- Check if there is pending output data and push the data toward downstream.
- Pulling events (from streaming thread) until there's at least
one pending "METransformNeedInput" event
- Then, push one data into MFT from streaming thread
- Check if there is pending "METransformHaveOutput" again.
If there is, push new output data to downstream
(unlikely there is pending output data at this moment)
Above flow was processed from upstream streaming thread. That means
even if there's available output data, it could be outputted later
when the next buffer is pushed from upstream streaming thread.
It would introduce at least one frame latency in case of live stream.
To reduce such latency, this commit modifies the flow to be fully
asynchronous like hardware MFT was designed and to be able to
output encoded data whenever it's available. More specifically,
IMFAsyncCallback object will be used for handling
"METransformNeedInput" and "METransformHaveOutput" events from
Media Foundation's internal thread, and new output data will be
also outputted from the Media Foundation's thread.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1520>
2020-08-18 18:19:26 +00:00
|
|
|
GST_VIDEO_ENCODER_STREAM_UNLOCK (self);
|
|
|
|
|
|
|
|
return S_OK;
|
|
|
|
}
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
typedef struct
|
|
|
|
{
|
|
|
|
guint profile;
|
|
|
|
const gchar *profile_str;
|
2022-01-27 16:45:30 +00:00
|
|
|
} GstMFVideoEncoderProfileMap;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_enum_internal (GstMFTransform * transform, GUID & subtype,
|
|
|
|
GstObject * d3d11_device, GstMFVideoEncoderDeviceCaps * device_caps,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstCaps ** sink_template, GstCaps ** src_template)
|
2020-12-26 11:39:07 +00:00
|
|
|
{
|
|
|
|
HRESULT hr;
|
|
|
|
MFT_REGISTER_TYPE_INFO *infos;
|
|
|
|
UINT32 info_size;
|
|
|
|
gint i;
|
2022-01-23 17:41:29 +00:00
|
|
|
GstCaps *src_caps = nullptr;
|
|
|
|
GstCaps *sink_caps = nullptr;
|
|
|
|
GstCaps *d3d11_caps = nullptr;
|
|
|
|
GValue *supported_formats = nullptr;
|
|
|
|
GValue *profiles = nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
gboolean have_I420 = FALSE;
|
2020-12-20 20:11:03 +00:00
|
|
|
gboolean have_NV12 = FALSE;
|
2021-01-20 19:41:44 +00:00
|
|
|
gboolean have_P010 = FALSE;
|
2021-10-01 11:27:28 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
2020-12-20 20:11:03 +00:00
|
|
|
gboolean d3d11_aware = FALSE;
|
2021-10-01 11:27:28 +00:00
|
|
|
#endif
|
2022-01-23 17:41:29 +00:00
|
|
|
gchar *device_name = nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
IMFActivate *activate;
|
|
|
|
IMFTransform *encoder;
|
|
|
|
ICodecAPI *codec_api;
|
2021-07-03 13:56:48 +00:00
|
|
|
ComPtr < IMFMediaType > out_type;
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderProfileMap h264_profile_map[] = {
|
2021-07-03 13:56:48 +00:00
|
|
|
{eAVEncH264VProfile_High, "high"},
|
|
|
|
{eAVEncH264VProfile_Main, "main"},
|
|
|
|
{eAVEncH264VProfile_Base, "baseline"},
|
2022-01-23 17:41:29 +00:00
|
|
|
{0, nullptr},
|
2020-12-26 11:39:07 +00:00
|
|
|
};
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderProfileMap hevc_profile_map[] = {
|
2021-07-03 13:56:48 +00:00
|
|
|
{eAVEncH265VProfile_Main_420_8, "main"},
|
|
|
|
{eAVEncH265VProfile_Main_420_10, "main-10"},
|
2022-01-23 17:41:29 +00:00
|
|
|
{0, nullptr},
|
2020-12-26 11:39:07 +00:00
|
|
|
};
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderProfileMap *profile_to_check = nullptr;
|
2021-10-15 14:18:41 +00:00
|
|
|
static const gchar *h264_caps_str =
|
2020-12-26 11:39:07 +00:00
|
|
|
"video/x-h264, stream-format=(string) byte-stream, alignment=(string) au";
|
2021-10-15 14:18:41 +00:00
|
|
|
static const gchar *hevc_caps_str =
|
2020-12-26 11:39:07 +00:00
|
|
|
"video/x-h265, stream-format=(string) byte-stream, alignment=(string) au";
|
2021-10-15 14:18:41 +00:00
|
|
|
static const gchar *vp9_caps_str = "video/x-vp9";
|
2022-01-23 17:41:29 +00:00
|
|
|
const gchar *codec_caps_str = nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
/* NOTE: depending on environment,
|
|
|
|
* some enumerated h/w MFT might not be usable (e.g., multiple GPU case) */
|
|
|
|
if (!gst_mf_transform_open (transform))
|
|
|
|
return;
|
|
|
|
|
|
|
|
activate = gst_mf_transform_get_activate_handle (transform);
|
|
|
|
if (!activate) {
|
|
|
|
GST_WARNING_OBJECT (transform, "No IMFActivate interface available");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
encoder = gst_mf_transform_get_transform_handle (transform);
|
|
|
|
if (!encoder) {
|
|
|
|
GST_WARNING_OBJECT (transform, "No IMFTransform interface available");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
codec_api = gst_mf_transform_get_codec_api_handle (transform);
|
|
|
|
if (!codec_api) {
|
|
|
|
GST_WARNING_OBJECT (transform, "No ICodecAPI interface available");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
g_object_get (transform, "device-name", &device_name, nullptr);
|
2020-12-26 11:39:07 +00:00
|
|
|
if (!device_name) {
|
|
|
|
GST_WARNING_OBJECT (transform, "Unknown device name");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
g_free (device_name);
|
|
|
|
|
|
|
|
hr = activate->GetAllocatedBlob (MFT_INPUT_TYPES_Attributes,
|
|
|
|
(UINT8 **) & infos, &info_size);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (i = 0; i < info_size / sizeof (MFT_REGISTER_TYPE_INFO); i++) {
|
|
|
|
GstVideoFormat format;
|
2021-07-15 07:44:18 +00:00
|
|
|
const GstVideoFormatInfo *format_info;
|
2020-12-26 11:39:07 +00:00
|
|
|
GValue val = G_VALUE_INIT;
|
|
|
|
|
|
|
|
format = gst_mf_video_subtype_to_video_format (&infos[i].guidSubtype);
|
|
|
|
if (format == GST_VIDEO_FORMAT_UNKNOWN)
|
|
|
|
continue;
|
|
|
|
|
2021-07-15 07:44:18 +00:00
|
|
|
format_info = gst_video_format_get_info (format);
|
|
|
|
if (GST_VIDEO_FORMAT_INFO_IS_RGB (format_info)) {
|
|
|
|
GST_DEBUG_OBJECT (transform, "Skip %s format",
|
|
|
|
GST_VIDEO_FORMAT_INFO_NAME (format_info));
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2020-12-26 11:39:07 +00:00
|
|
|
if (!supported_formats) {
|
|
|
|
supported_formats = g_new0 (GValue, 1);
|
|
|
|
g_value_init (supported_formats, GST_TYPE_LIST);
|
|
|
|
}
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
switch (format) {
|
2021-07-03 13:56:48 +00:00
|
|
|
/* media foundation has duplicated formats IYUV and I420 */
|
2020-12-20 20:11:03 +00:00
|
|
|
case GST_VIDEO_FORMAT_I420:
|
|
|
|
if (have_I420)
|
|
|
|
continue;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
have_I420 = TRUE;
|
|
|
|
break;
|
|
|
|
case GST_VIDEO_FORMAT_NV12:
|
|
|
|
have_NV12 = TRUE;
|
|
|
|
break;
|
2021-01-20 19:41:44 +00:00
|
|
|
case GST_VIDEO_FORMAT_P010_10LE:
|
|
|
|
have_P010 = TRUE;
|
|
|
|
break;
|
2020-12-20 20:11:03 +00:00
|
|
|
default:
|
|
|
|
break;
|
2020-12-26 11:39:07 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
g_value_init (&val, G_TYPE_STRING);
|
|
|
|
g_value_set_static_string (&val, gst_video_format_to_string (format));
|
|
|
|
gst_value_list_append_and_take_value (supported_formats, &val);
|
|
|
|
}
|
|
|
|
CoTaskMemFree (infos);
|
|
|
|
|
|
|
|
if (!supported_formats) {
|
|
|
|
GST_WARNING_OBJECT (transform, "Couldn't figure out supported format");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (IsEqualGUID (MFVideoFormat_H264, subtype)) {
|
|
|
|
profile_to_check = h264_profile_map;
|
|
|
|
codec_caps_str = h264_caps_str;
|
|
|
|
} else if (IsEqualGUID (MFVideoFormat_HEVC, subtype)) {
|
|
|
|
profile_to_check = hevc_profile_map;
|
|
|
|
codec_caps_str = hevc_caps_str;
|
|
|
|
} else if (IsEqualGUID (MFVideoFormat_VP90, subtype)) {
|
|
|
|
codec_caps_str = vp9_caps_str;
|
|
|
|
} else {
|
|
|
|
g_assert_not_reached ();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (profile_to_check) {
|
|
|
|
hr = MFCreateMediaType (&out_type);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
hr = out_type->SetGUID (MF_MT_MAJOR_TYPE, MFMediaType_Video);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
hr = out_type->SetGUID (MF_MT_SUBTYPE, subtype);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
hr = out_type->SetUINT32 (MF_MT_AVG_BITRATE, 2048000);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
hr = MFSetAttributeRatio (out_type.Get (), MF_MT_FRAME_RATE, 30, 1);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
2021-07-03 13:56:48 +00:00
|
|
|
hr = out_type->SetUINT32 (MF_MT_INTERLACE_MODE,
|
|
|
|
MFVideoInterlace_Progressive);
|
2020-12-26 11:39:07 +00:00
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
hr = MFSetAttributeSize (out_type.Get (), MF_MT_FRAME_SIZE, 1920, 1080);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
i = 0;
|
|
|
|
do {
|
|
|
|
GValue profile_val = G_VALUE_INIT;
|
|
|
|
guint mf_profile = profile_to_check[i].profile;
|
|
|
|
const gchar *profile_str = profile_to_check[i].profile_str;
|
|
|
|
|
|
|
|
i++;
|
|
|
|
|
|
|
|
if (mf_profile == 0)
|
|
|
|
break;
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
g_assert (profile_str != nullptr);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
hr = out_type->SetUINT32 (MF_MT_MPEG2_PROFILE, mf_profile);
|
|
|
|
if (!gst_mf_result (hr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!gst_mf_transform_set_output_type (transform, out_type.Get ()))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!profiles) {
|
|
|
|
profiles = g_new0 (GValue, 1);
|
|
|
|
g_value_init (profiles, GST_TYPE_LIST);
|
|
|
|
}
|
|
|
|
|
2021-06-21 17:34:18 +00:00
|
|
|
/* Add "constrained-baseline" in addition to "baseline" */
|
|
|
|
if (profile_str == "baseline") {
|
|
|
|
g_value_init (&profile_val, G_TYPE_STRING);
|
|
|
|
g_value_set_static_string (&profile_val, "constrained-baseline");
|
|
|
|
gst_value_list_append_and_take_value (profiles, &profile_val);
|
|
|
|
}
|
|
|
|
|
2020-12-26 11:39:07 +00:00
|
|
|
g_value_init (&profile_val, G_TYPE_STRING);
|
|
|
|
g_value_set_static_string (&profile_val, profile_str);
|
|
|
|
gst_value_list_append_and_take_value (profiles, &profile_val);
|
|
|
|
} while (1);
|
|
|
|
|
|
|
|
if (!profiles) {
|
|
|
|
GST_WARNING_OBJECT (transform, "Couldn't query supported profile");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
src_caps = gst_caps_from_string (codec_caps_str);
|
|
|
|
if (profiles) {
|
|
|
|
gst_caps_set_value (src_caps, "profile", profiles);
|
|
|
|
g_value_unset (profiles);
|
|
|
|
g_free (profiles);
|
|
|
|
}
|
|
|
|
|
|
|
|
sink_caps = gst_caps_new_empty_simple ("video/x-raw");
|
|
|
|
/* FIXME: don't hardcode max resolution, but MF doesn't provide
|
|
|
|
* API for querying supported max resolution... */
|
|
|
|
gst_caps_set_simple (sink_caps,
|
|
|
|
"width", GST_TYPE_INT_RANGE, 64, 8192,
|
2022-01-23 17:41:29 +00:00
|
|
|
"height", GST_TYPE_INT_RANGE, 64, 8192, nullptr);
|
2020-12-26 11:39:07 +00:00
|
|
|
gst_caps_set_simple (src_caps,
|
|
|
|
"width", GST_TYPE_INT_RANGE, 64, 8192,
|
2022-01-23 17:41:29 +00:00
|
|
|
"height", GST_TYPE_INT_RANGE, 64, 8192, nullptr);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
#if GST_MF_HAVE_D3D11
|
|
|
|
/* Check whether this MFT can support D3D11 */
|
2021-01-20 19:41:44 +00:00
|
|
|
if (d3d11_device && (have_NV12 || have_P010)) {
|
2022-01-23 17:41:29 +00:00
|
|
|
g_object_get (transform, "d3d11-aware", &d3d11_aware, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
GST_DEBUG_OBJECT (transform, "d3d11 aware %d", d3d11_aware);
|
|
|
|
}
|
|
|
|
|
2021-01-20 19:41:44 +00:00
|
|
|
if (d3d11_device && (have_NV12 || have_P010) && d3d11_aware) {
|
2021-10-08 14:07:32 +00:00
|
|
|
gint64 adapter_luid = 0;
|
2021-01-20 19:41:44 +00:00
|
|
|
GValue d3d11_formats = G_VALUE_INIT;
|
2020-12-20 20:11:03 +00:00
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
g_object_get (d3d11_device, "adapter-luid", &adapter_luid, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
|
|
|
|
d3d11_caps = gst_caps_copy (sink_caps);
|
|
|
|
|
2021-01-20 19:41:44 +00:00
|
|
|
g_value_init (&d3d11_formats, GST_TYPE_LIST);
|
|
|
|
if (have_NV12) {
|
|
|
|
GValue val = G_VALUE_INIT;
|
|
|
|
g_value_init (&val, G_TYPE_STRING);
|
|
|
|
g_value_set_static_string (&val, "NV12");
|
|
|
|
gst_value_list_append_and_take_value (&d3d11_formats, &val);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (have_P010) {
|
|
|
|
GValue val = G_VALUE_INIT;
|
|
|
|
g_value_init (&val, G_TYPE_STRING);
|
|
|
|
g_value_set_static_string (&val, "P010_10LE");
|
|
|
|
gst_value_list_append_and_take_value (&d3d11_formats, &val);
|
|
|
|
}
|
|
|
|
|
|
|
|
gst_caps_set_value (d3d11_caps, "format", &d3d11_formats);
|
|
|
|
g_value_unset (&d3d11_formats);
|
2020-12-20 20:11:03 +00:00
|
|
|
gst_caps_set_features_simple (d3d11_caps,
|
|
|
|
gst_caps_features_from_string (GST_CAPS_FEATURE_MEMORY_D3D11_MEMORY));
|
|
|
|
device_caps->d3d11_aware = TRUE;
|
2021-10-08 14:07:32 +00:00
|
|
|
device_caps->adapter_luid = adapter_luid;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
gst_caps_set_value (sink_caps, "format", supported_formats);
|
|
|
|
g_value_unset (supported_formats);
|
|
|
|
g_free (supported_formats);
|
|
|
|
|
|
|
|
if (d3d11_caps)
|
|
|
|
gst_caps_append (sink_caps, d3d11_caps);
|
|
|
|
|
2020-12-26 11:39:07 +00:00
|
|
|
*sink_template = sink_caps;
|
|
|
|
*src_template = src_caps;
|
|
|
|
|
|
|
|
#define CHECK_DEVICE_CAPS(codec_obj,api,val) \
|
|
|
|
if (SUCCEEDED((codec_obj)->IsSupported(&(api)))) {\
|
|
|
|
(device_caps)->val = TRUE; \
|
|
|
|
}
|
|
|
|
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncCommonRateControlMode, rc_mode);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncCommonQuality, quality);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncAdaptiveMode, adaptive_mode);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncCommonBufferSize, buffer_size);
|
2022-03-26 14:48:46 +00:00
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncCommonMeanBitRate, mean_bitrate);
|
2020-12-26 11:39:07 +00:00
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncCommonMaxBitRate, max_bitrate);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api,
|
|
|
|
CODECAPI_AVEncCommonQualityVsSpeed, quality_vs_speed);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncH264CABACEnable, cabac);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncH264SPSID, sps_id);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncH264PPSID, pps_id);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncMPVDefaultBPictureCount, bframes);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncMPVGOPSize, gop_size);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncNumWorkerThreads, threads);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncVideoContentType, content_type);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncVideoEncodeQP, qp);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api,
|
|
|
|
CODECAPI_AVEncVideoForceKeyFrame, force_keyframe);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVLowLatencyMode, low_latency);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncVideoMinQP, min_qp);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncVideoMaxQP, max_qp);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api,
|
|
|
|
CODECAPI_AVEncVideoEncodeFrameTypeQP, frame_type_qp);
|
|
|
|
CHECK_DEVICE_CAPS (codec_api, CODECAPI_AVEncVideoMaxNumRefFrame, max_num_ref);
|
|
|
|
if (device_caps->max_num_ref) {
|
|
|
|
VARIANT min;
|
|
|
|
VARIANT max;
|
|
|
|
VARIANT step;
|
|
|
|
|
|
|
|
hr = codec_api->GetParameterRange (&CODECAPI_AVEncVideoMaxNumRefFrame,
|
|
|
|
&min, &max, &step);
|
|
|
|
if (SUCCEEDED (hr)) {
|
|
|
|
device_caps->max_num_ref_high = max.uiVal;
|
|
|
|
device_caps->max_num_ref_low = min.uiVal;
|
|
|
|
VariantClear (&min);
|
|
|
|
VariantClear (&max);
|
|
|
|
VariantClear (&step);
|
|
|
|
} else {
|
|
|
|
device_caps->max_num_ref = FALSE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#undef CHECK_DEVICE_CAPS
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
static GstMFTransform *
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_enum (guint enum_flags, GUID * subtype, guint device_index,
|
|
|
|
GstMFVideoEncoderDeviceCaps * device_caps, GstObject * d3d11_device,
|
2020-12-20 20:11:03 +00:00
|
|
|
GstCaps ** sink_template, GstCaps ** src_template)
|
2020-12-26 11:39:07 +00:00
|
|
|
{
|
|
|
|
GstMFTransformEnumParams enum_params = { 0, };
|
|
|
|
MFT_REGISTER_TYPE_INFO output_type;
|
|
|
|
GstMFTransform *transform;
|
2020-12-20 20:11:03 +00:00
|
|
|
gint64 adapter_luid = 0;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
*sink_template = nullptr;
|
|
|
|
*src_template = nullptr;
|
2022-01-27 16:45:30 +00:00
|
|
|
memset (device_caps, 0, sizeof (GstMFVideoEncoderDeviceCaps));
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
if (!IsEqualGUID (MFVideoFormat_H264, *subtype) &&
|
|
|
|
!IsEqualGUID (MFVideoFormat_HEVC, *subtype) &&
|
|
|
|
!IsEqualGUID (MFVideoFormat_VP90, *subtype)) {
|
|
|
|
GST_ERROR ("Unknown subtype GUID");
|
|
|
|
|
2022-01-23 17:41:29 +00:00
|
|
|
return nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
}
|
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
if (d3d11_device) {
|
2022-01-23 17:41:29 +00:00
|
|
|
g_object_get (d3d11_device, "adapter-luid", &adapter_luid, nullptr);
|
2020-12-20 20:11:03 +00:00
|
|
|
if (!adapter_luid) {
|
|
|
|
GST_ERROR ("Couldn't get adapter LUID");
|
2022-01-23 17:41:29 +00:00
|
|
|
return nullptr;
|
2020-12-20 20:11:03 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-12-26 11:39:07 +00:00
|
|
|
output_type.guidMajorType = MFMediaType_Video;
|
|
|
|
output_type.guidSubtype = *subtype;
|
|
|
|
|
|
|
|
enum_params.category = MFT_CATEGORY_VIDEO_ENCODER;
|
|
|
|
enum_params.output_typeinfo = &output_type;
|
|
|
|
enum_params.device_index = device_index;
|
|
|
|
enum_params.enum_flags = enum_flags;
|
2020-12-20 20:11:03 +00:00
|
|
|
enum_params.adapter_luid = adapter_luid;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
transform = gst_mf_transform_new (&enum_params);
|
|
|
|
if (!transform)
|
2022-01-23 17:41:29 +00:00
|
|
|
return nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_enum_internal (transform, output_type.guidSubtype,
|
2020-12-20 20:11:03 +00:00
|
|
|
d3d11_device, device_caps, sink_template, src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
return transform;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_register_internal (GstPlugin * plugin, guint rank,
|
2020-12-26 11:39:07 +00:00
|
|
|
GUID * subtype, GTypeInfo * type_info,
|
2022-01-27 16:45:30 +00:00
|
|
|
const GstMFVideoEncoderDeviceCaps * device_caps,
|
2020-12-26 11:39:07 +00:00
|
|
|
guint32 enum_flags, guint device_index, GstMFTransform * transform,
|
|
|
|
GstCaps * sink_caps, GstCaps * src_caps)
|
|
|
|
{
|
|
|
|
GType type;
|
|
|
|
GTypeInfo local_type_info;
|
|
|
|
gchar *type_name;
|
|
|
|
gchar *feature_name;
|
|
|
|
gint i;
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderClassData *cdata;
|
2020-12-26 11:39:07 +00:00
|
|
|
gboolean is_default = TRUE;
|
2022-01-23 17:41:29 +00:00
|
|
|
gchar *device_name = nullptr;
|
|
|
|
const gchar *type_name_prefix = nullptr;
|
|
|
|
const gchar *feature_name_prefix = nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
if (IsEqualGUID (MFVideoFormat_H264, *subtype)) {
|
|
|
|
type_name_prefix = "H264";
|
|
|
|
feature_name_prefix = "h264";
|
|
|
|
} else if (IsEqualGUID (MFVideoFormat_HEVC, *subtype)) {
|
|
|
|
type_name_prefix = "H265";
|
|
|
|
feature_name_prefix = "h265";
|
|
|
|
} else if (IsEqualGUID (MFVideoFormat_VP90, *subtype)) {
|
|
|
|
type_name_prefix = "VP9";
|
|
|
|
feature_name_prefix = "vp9";
|
|
|
|
} else {
|
|
|
|
g_assert_not_reached ();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Must be checked already */
|
2022-01-23 17:41:29 +00:00
|
|
|
g_object_get (transform, "device-name", &device_name, nullptr);
|
|
|
|
g_assert (device_name != nullptr);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
cdata = g_new0 (GstMFVideoEncoderClassData, 1);
|
2020-12-26 11:39:07 +00:00
|
|
|
cdata->sink_caps = gst_caps_copy (sink_caps);
|
|
|
|
cdata->src_caps = gst_caps_copy (src_caps);
|
|
|
|
cdata->device_name = device_name;
|
|
|
|
cdata->device_caps = *device_caps;
|
|
|
|
cdata->enum_flags = enum_flags;
|
|
|
|
cdata->device_index = device_index;
|
|
|
|
|
|
|
|
local_type_info = *type_info;
|
|
|
|
local_type_info.class_data = cdata;
|
|
|
|
|
|
|
|
GST_MINI_OBJECT_FLAG_SET (cdata->sink_caps,
|
|
|
|
GST_MINI_OBJECT_FLAG_MAY_BE_LEAKED);
|
|
|
|
GST_MINI_OBJECT_FLAG_SET (cdata->src_caps,
|
|
|
|
GST_MINI_OBJECT_FLAG_MAY_BE_LEAKED);
|
|
|
|
|
|
|
|
type_name = g_strdup_printf ("GstMF%sEnc", type_name_prefix);
|
|
|
|
feature_name = g_strdup_printf ("mf%senc", feature_name_prefix);
|
|
|
|
|
|
|
|
i = 1;
|
|
|
|
while (g_type_from_name (type_name) != 0) {
|
|
|
|
g_free (type_name);
|
|
|
|
g_free (feature_name);
|
|
|
|
type_name = g_strdup_printf ("GstMF%sDevice%dEnc", type_name_prefix, i);
|
|
|
|
feature_name = g_strdup_printf ("mf%sdevice%denc", feature_name_prefix, i);
|
|
|
|
is_default = FALSE;
|
|
|
|
i++;
|
|
|
|
}
|
|
|
|
|
|
|
|
cdata->is_default = is_default;
|
|
|
|
|
|
|
|
type =
|
2022-01-27 16:45:30 +00:00
|
|
|
g_type_register_static (GST_TYPE_MF_VIDEO_ENCODER, type_name,
|
2020-12-26 11:39:07 +00:00
|
|
|
&local_type_info, (GTypeFlags) 0);
|
|
|
|
|
|
|
|
/* make lower rank than default device */
|
|
|
|
if (rank > 0 && !is_default)
|
|
|
|
rank--;
|
|
|
|
|
2021-11-17 13:46:17 +00:00
|
|
|
if (!is_default)
|
|
|
|
gst_element_type_set_skip_documentation (type);
|
|
|
|
|
2020-12-26 11:39:07 +00:00
|
|
|
if (!gst_element_register (plugin, feature_name, rank, type))
|
|
|
|
GST_WARNING ("Failed to register plugin '%s'", type_name);
|
|
|
|
|
|
|
|
g_free (type_name);
|
|
|
|
g_free (feature_name);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_register (GstPlugin * plugin, guint rank, GUID * subtype,
|
2020-12-20 20:11:03 +00:00
|
|
|
GTypeInfo * type_info, GList * d3d11_device)
|
2020-12-26 11:39:07 +00:00
|
|
|
{
|
2022-01-23 17:41:29 +00:00
|
|
|
GstMFTransform *transform = nullptr;
|
|
|
|
GstCaps *sink_template = nullptr;
|
|
|
|
GstCaps *src_template = nullptr;
|
2020-12-26 11:39:07 +00:00
|
|
|
guint enum_flags;
|
2022-01-27 16:45:30 +00:00
|
|
|
GstMFVideoEncoderDeviceCaps device_caps;
|
2020-12-26 11:39:07 +00:00
|
|
|
guint i;
|
|
|
|
|
|
|
|
/* register hardware encoders first */
|
|
|
|
enum_flags = (MFT_ENUM_FLAG_HARDWARE | MFT_ENUM_FLAG_ASYNCMFT |
|
2022-03-26 13:24:27 +00:00
|
|
|
MFT_ENUM_FLAG_SORTANDFILTER_APPROVED_ONLY);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
if (d3d11_device) {
|
|
|
|
GList *iter;
|
|
|
|
for (iter = d3d11_device; iter; iter = g_list_next (iter)) {
|
|
|
|
GstObject *device = (GstObject *) iter->data;
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
transform =
|
|
|
|
gst_mf_video_encoder_enum (enum_flags, subtype, 0, &device_caps,
|
2020-12-20 20:11:03 +00:00
|
|
|
device, &sink_template, &src_template);
|
|
|
|
|
|
|
|
/* No more MFT to enumerate */
|
|
|
|
if (!transform)
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* Failed to open MFT */
|
|
|
|
if (!sink_template) {
|
|
|
|
gst_clear_object (&transform);
|
|
|
|
continue;
|
|
|
|
}
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_register_internal (plugin, rank, subtype, type_info,
|
2020-12-20 20:11:03 +00:00
|
|
|
&device_caps, enum_flags, 0, transform, sink_template, src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
gst_clear_object (&transform);
|
2020-12-20 20:11:03 +00:00
|
|
|
gst_clear_caps (&sink_template);
|
|
|
|
gst_clear_caps (&src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
}
|
2020-12-20 20:11:03 +00:00
|
|
|
} else {
|
|
|
|
/* AMD seems to be able to support up to 12 GPUs */
|
|
|
|
for (i = 0; i < 12; i++) {
|
2022-01-27 16:45:30 +00:00
|
|
|
transform =
|
|
|
|
gst_mf_video_encoder_enum (enum_flags, subtype, i, &device_caps,
|
2022-01-23 17:41:29 +00:00
|
|
|
nullptr, &sink_template, &src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2020-12-20 20:11:03 +00:00
|
|
|
/* No more MFT to enumerate */
|
|
|
|
if (!transform)
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* Failed to open MFT */
|
|
|
|
if (!sink_template) {
|
|
|
|
gst_clear_object (&transform);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_register_internal (plugin, rank, subtype, type_info,
|
2020-12-20 20:11:03 +00:00
|
|
|
&device_caps, enum_flags, i, transform, sink_template, src_template);
|
|
|
|
gst_clear_object (&transform);
|
|
|
|
gst_clear_caps (&sink_template);
|
|
|
|
gst_clear_caps (&src_template);
|
|
|
|
}
|
2020-12-26 11:39:07 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* register software encoders */
|
|
|
|
enum_flags = (MFT_ENUM_FLAG_SYNCMFT |
|
2022-03-26 13:24:27 +00:00
|
|
|
MFT_ENUM_FLAG_SORTANDFILTER_APPROVED_ONLY);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
transform = gst_mf_video_encoder_enum (enum_flags, subtype, 0, &device_caps,
|
2022-01-23 17:41:29 +00:00
|
|
|
nullptr, &sink_template, &src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
if (!transform)
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
if (!sink_template)
|
|
|
|
goto done;
|
|
|
|
|
2022-01-27 16:45:30 +00:00
|
|
|
gst_mf_video_encoder_register_internal (plugin, rank, subtype, type_info,
|
2020-12-26 18:43:11 +00:00
|
|
|
&device_caps, enum_flags, 0, transform, sink_template, src_template);
|
2020-12-26 11:39:07 +00:00
|
|
|
|
|
|
|
done:
|
|
|
|
gst_clear_object (&transform);
|
|
|
|
gst_clear_caps (&sink_template);
|
|
|
|
gst_clear_caps (&src_template);
|
2021-07-03 13:56:48 +00:00
|
|
|
}
|