Introduce GST_USE_NV_STATELESS_CODEC environment to allow user to select
primary nvidia decoder implementation. In case the environment
GST_USE_NV_STATELESS_CODEC=h264 was set, old nvidia h264 decoder element
will not be registered. Instead, both nvh264dec and nvh264sldec
factory name will create gstcodecs based new nvidia h264 decoder element.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1198>
Introduce GstH264Decoder based Nvidia H.264 decoder element.
Similar the element factory name of to v4l2 stateless codec,
this element can be configured with factory name "gstnvh264sldec".
Note that "sl" in the name stands for "stateless"
For now, existing nvh264dec covers more profile and formats
(e.g., interlaced stream) than this implementation.
However, this implementation allows us to control lower level
parameters such as decoded picture buffer management and therefore
we can get a chance to improve performance in terms of latency.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1198>
Too many decode surface would waste GPU memory. Also it seems to be
introducing additional latency depending on stream. Since nvcodec
sdk version 9.0, CUVID parser API has been providing the minimum
required number of surface. By using it, we can save GPU memory
and reduce possible latency.
The class data with the caps in it will be leaked if the element is
registered but never instantiated. There is no way around this. Mark
the caps as such so that the leaks tracer does not warn about it.
This is the same as pad template caps getting leaked, which are also
marked as may-be-leaked. These objects are initialized exactly once,
and are 'global' data.
We've been using NvEncodeAPICreateInstance method to find the supported API
version, but that seems to be insufficient since there is a case
where plugin failed in opening encoding session even if NvEncodeAPICreateInstance
succeeded. Asking driver about the version would be the most certain way.
Setting the CUVID_PKT_DISCONTINUITY implies clearing any past information
about the stream in the decoder. The GStreamer discont flag is used for
discontinuity caused by a seek, for first buffer and if a buffer was
dropped. In the first two cases, the parsers and demuxers should ensure we
start from a synchronization point, so it's unlikely that delta will be
matched against the wrong state.
For packet lost, the discontinuity flag will prevent the decoder from doing
any concealment, with a result that ca be much worst visually, or freeze the
playback until an IDR is met. It's better to let the decoder handle that for
us.
Removing this flag, also workaround a but in NVidia parser that makes it
ignore our ENDOFFRAME flag and increase the latency by one frame.
This sets the CUVID_PKT_ENDOFPICTURE flag in order to inform the decoder that
we have a complete picture. This should remove one frame latency otherwise
introduce by NVidia parser.
We weren't using the correct calling convention when calling CUDA and
CUVID APIs. `CUDAAPI` is `__stdcall` on Windows. This was working fine
on x64 because `__stdcall` is ignored and there's no special calling
convention. However, on x86, we need to use `__stdcall`.
Hard-coded 16x16 resolution is likely to differ from the device's support
in most cases. If we can use NV_ENC_CAPS_WIDTH_MIN and NV_ENC_CAPS_HEIGHT_MIN,
update pad template with returned value.
GstNvBaseEnc::n_bufs was set from the previous encoding session
but it wasn't cleared after stop. That might result to invalid memory
access at the next start (no encoded data) and then stop sequence.
Instead of defining a variable for array length, use GArray::len
directly to avoid such confusion.
If the last flow was not GST_FLOW_OK, the encoding thread is not running
and there is nothing to pop from GAsyncQueue (this causes deadlock).
To prevent deadlock, just return the handle_frame without further encoding
process if the last flow was not GST_FLOW_OK. Note that the last flow
will be cleared per FLUSH_STOP and STREAM_START event.
The hard-coded upper bound 32 (or 48 depending on resolution) might waste
GPU memory and high resolution encoding causes OUT-OF-MEMORY allocation error
quite easily. This commit calculates the number of required pre-allocated
device memory based on encoding options and it can reduce the amount of device memory
used by nvenc.
NVDEC driver always uses input timestamp without adjustment
even if bframe encoding was enabled.
So DTS can be larger than PTS when bframe was enabled.
To ensure PTS >= DTS, we should adjust the timestamp manually
based on the PTS difference between the first
encoded frame and the second one. That's also the maximum PTS/DTS
difference.
To support rc-lookahead and bframe encoding, nvenc needs one more
staging queue, because NvEncEncodePicture can return NV_ENC_ERR_NEED_MORE_INPUT
but which was not considered so far.
As documented by NVENC programming guide, pending buffers should wait
other inputs until NvEncEncodePicture returns success.
New encoding flow is
- Submit raw picture buffer to encoder with NvEncEncodePicture
- The submitted input/output buffer pair will be queued to pending_queue
- If NvEncEncodePicture returned success, then move all pair in pending_queue
to final stage
- Otherwise, wait more input raw pictures.
Another change is dropping NV_ENC_LOCK_INPUT_BUFFER usage.
So now nvenc always uses CUDA memory input buffer. As a result,
both opengl and system memory handling are unified.
* The number of iteration is always one so the iteration is useless
and that makes code complicated.
* Also defining named structure can code mroe readable.
* g_free is null safe
New rate-control modes are introduced (if device can support)
* cbr-ld-hr: CBR low-delay high quality
* cbr-hq: CBR high quality
* vbr-hq: VBR high quality
Also, various configurable rate-control related properties are added.
Introducing new dynamic class between GstNvBaseEncClass and
each subclass to be able to access device specific properties and
capabilities from each subclass implementation side.
Do not restrict allowed maximum resolution depending on the
initial resolution. If new resolution is larger than previous one,
just re-init encode session.
Due to uncleared last flow, decoding after seek was never possible
(last_ret == GST_FLOW_FLUSHING).
nvdec dose not need to keep track of the previous flow return,
and actually the interest is data/even flow of the current handle_frame().
Implementing ::negotiate() method to support runtime output format
change. If downstream was reconfigured, baseclass will invoke
::negotiate() method, and nvdec should update output memory
type depending on downstream caps.
Input stream might be silently changed without ::set_format() call.
Since nvdec has internal parser, nvdec element can figure out the format change
by itself.
Register openGL resource only once per memory. Also if upstream
provides the registered information, reuse the information
instead of doing it again. This can improve performance dramatically
depending on system since the resource registration might cause
high overhead.
Introduce GstCudaGraphicsResource structure to represent registered
CUDA graphics resources and to enable sharing the information among
nvdec and nvenc. This structure can reduce the number of resource
registration which cause high overhead.
For openGL interoperability, nvdec uses cuGraphicsGLRegisterImage API
which is to register openGL texture image.
Meanwhile nvenc uses cuGraphicsGLRegisterBuffer API to registure openGL buffer object.
That means two kinds of graphics resources are registered per memory
when nvdec/nvenc are configured at the same time.
The graphics resource registration brings possibly high overhead
so the registration should be performed only once per resource
from optimization point of view.
gst_query_get_n_allocation_pools > 0 does not guarantee that
the N th internal array has GstBufferPool object. So users should
check the returned GstBufferPool object from
gst_query_parse_nth_allocation_pool.
Async CUDA operation with default stream (NULL CUstream) is not much
beneficial than blocking operation since all CUDA operations which belong
to the CUDA context will be synchronized with the default stream's operation.
Note that CUDA stream will share all resources of the corresponding CUDA context
but which can help parallel operation similar to the relation between thread and process
The internal decoding state must be GST_NVDEC_STATE_PARSE before
calling CuvidParseVideoData(). Otherwise, nvdec will be confused
on decode callback as if the frame is decoding only frame and
the input timestamp of corresponding frame will be ignored.
Eventually one decoded frame will have non-increased PTS.
The destroy callback can be called just before the fìnalization of
GstMiniObject. So the nvdec object might be destroyed already.
Instead, store the GstCudaContext with increased ref to safely
unregister the CUDA resource.
YV12 format is supported by Nvidia NVENC without manual conversion.
So nvenc is exposing YV12 format at sinkpad template but there is some
missing point around uploading the memory to GPU.
The gst_cuda_result macro function is more helpful for debugging
than previous cuda_OK because gst_cuda_result prints the function
and line number. If the CUDA API return was not CUDA_SUCCESS,
gst_cuda_result will print WARNING level debug message with
error name, error text strings.
... and drop CUvideoctxlock usage. The CUvideoctxlock basically
has the identical role of cuda context push/pop but nvdec specific
way. Since we can share the CUDA context among encoders and decoders,
use CUDA context directly for accessing GPU API.
... and add support CUDA context sharing similar to glcontext sharing.
Multiple CUDA context per GPU is not the best practice. The context
sharing method is very similar to that of glcontext. The difference
is that there can be multiple context object on a pipeline since
the CUDA context is created per GPU id. For example, a pipeline
has nvh264dec (uses GPU #0) and nvh264device0dec (uses GPU #1),
then two CUDA context will propagated to all pipeline.
New object and helper functions can remove duplicated code
from nvenc/nvdec. Also this is prework for CUDA device context sharing
among nvdec(s)/nvenc(s).
During GstVideoInfo conversion from GstCaps, interlace-mode is
inferred to progressive so unspecified interlace-mode should not cause any
negotiation issue. Simly set GST_PAD_FLAG_ACCEPT_INTERSECT flag
on sinkpad to fix issue.
Encoded bitstream might not have valid framerate. If upstream
provided non-variable-framerate (i.e., fps_n > 0 and fps_d > 0)
use upstream framerate instead of parsed one.
Encoding thread is terminated without any notification so
upstream streaming thread is locked because there is nothing
to pop from GAsyncQueue. If downstream returns error,
we need put SHUTDOWN_COOKIE to GAsyncQueue for chain function
can wakeup.
By adding system memory support for nvdec, both en/decoder
in the nvcodec plugin are able to be usable regardless of
OpenGL dependency. Besides, the direct use of system memory
might have less overhead than OpenGL memory depending on use cases.
(e.g., transcoding using S/W encoder)
Any plugin which returned FALSE from plugin_init will be blacklisted
so the plugin will be unusable even if an user install required runtime
dependency next time. So that's the reason why nvcodec returns TRUE always.
This commit is to remove possible misreading code.
Since we build nvcodec plugin without external CUDA dependency,
CUDA and en/decoder library loading failure can be natural behavior.
Emit error only when the module was opend but required symbols are missing.
This commit includes h265 main-10 profile support if the device can
decode it.
Note that since h264 10bits decoding is not supported by nvidia GPU for now,
the additional code path for h264 high-10 profile is a preparation for
the future Nvidia's enhancement.
GstVideoDecoder::drain/flush can be called at very initial state
with stream-start and flush-stop event, respectively.
Draning with NULL CUvideoparser seems to unsafe and that eventually
failed to handle it.
... and add our stub cuda header.
Newly introduced stub cuda.h file is defining minimal types in order to
build nvcodec plugin without system installed CUDA toolkit dependency.
This will make cross-compile possible.
* By this commit, if there are more than one device,
nvenc element factory will be created per
device like nvh264device{device-id}enc and nvh265device{device-id}enc
in addition to nvh264enc and nvh265enc, so that the element factory
can expose the exact capability of the device for the codec.
* Each element factory will have fixed cuda-device-id
which is determined during plugin initialization
depending on the capability of corresponding device.
(e.g., when only the second device can encode h265 among two GPU,
then nvh265enc will choose "1" (zero-based numbering)
as it's target cuda-device-id. As we have element factory
per GPU device, "cuda-device-id" property is changed to read-only.
* nvh265enc gains ability to encoding
4:4:4 8bits, 4:2:0 10 bits formats and up to 8K resolution
depending on device capability.
Additionally, I420 GLMemory input is supported by nvenc.
Only the default device has been used by NVDEC so far.
This commit make it possible to use registered device id.
To simplify device id selection, GstNvDecCudaContext usage is removed.
By this commit, each codec has its own element factory so the
nvdec element factory is removed. Also, if there are more than one device,
additional nvdec element factory will be created per
device like nvh264device{device-id}dec, so that the element factory
can expose the exact capability of the device for the codec.
Callbacks of CUvideoparser is called on the streaming thread.
So the use of async queue has no benefit.
Make control flow straightforward instead of long while/switch loop.
