Phase 1: NVENC через dlopen + источник через cuframes_subscriber

Скелет проекта cuframes-composer (LGPL-2.1+) и MVP кодирования
одного источника в файл H.264.

Что включает Phase 1:

- LICENSE (LGPL-2.1+), README с поэтапным планом, корневой CMake
- Подмодули: cuframes v0.4 (pinned), nv-codec-headers (n12.2.72.0)
- include/cuframes_composer/source.h — публичный API источника
  с явной машиной состояний (DISCONNECTED → CONNECTING → ACTIVE →
  STALE → DEAD) и snapshot-паттерном для чтения без блокировки
- include/cuframes_composer/nvenc.h — публичный API кодировщика
  на CUdeviceptr-вход (zero-copy через VMM-mapped указатели)
- src/nvenc_loader.{h,c} — dlopen libnvidia-encode.so.1 и инициализация
  таблицы функций NVENC через NvEncodeAPICreateInstance. Идёт через
  pthread_once. Сделано отдельно чтобы держать LGPL-совместимость:
  проприетарный SDK не статически линкуется
- src/nvenc.c — обвязка над NVENC: open session, init encoder, кеш
  registered resources, encode/lock/unlock, flush с EOS, поддержка
  H.264 CBR low-latency, preset GUID p1/p4/p7
- src/source.c — обвязка над cuframes_subscriber c фоновым потоком,
  exponential backoff reconnect (1с → 30с), и переходами по таймаутам
  для stale/dead-детекта
- examples/simple_record — smoke-test программа: подписка на cuframes,
  кодирование, запись в .h264 файл, корректное завершение по SIGINT
This commit is contained in:
2026-06-03 04:28:33 +01:00
commit ba68550f4c
16 changed files with 2051 additions and 0 deletions
+31
View File
@@ -0,0 +1,31 @@
# Build outputs
build/
build-*/
cmake-build-*/
*.o
*.so
*.a
# IDE
.vscode/
.idea/
.cache/
compile_commands.json
# Editor backups
*~
.*.swp
# Local config
.env
.env.local
# Coverage / profiling
*.gcda
*.gcno
*.gcov
callgrind.out.*
perf.data*
# Generated docs
docs/_build/
+6
View File
@@ -0,0 +1,6 @@
[submodule "third_party/cuframes"]
path = third_party/cuframes
url = /home/claude/projects/cuframes
[submodule "third_party/nv-codec-headers"]
path = third_party/nv-codec-headers
url = https://github.com/FFmpeg/nv-codec-headers.git
+75
View File
@@ -0,0 +1,75 @@
cmake_minimum_required(VERSION 3.20)
project(cuframes-composer
VERSION 0.1.0
DESCRIPTION "Multi-source video grid composer на CUDA + NVENC + RTSP"
LANGUAGES C
)
set(CMAKE_C_STANDARD 11)
set(CMAKE_C_STANDARD_REQUIRED ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
if(NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE Release)
endif()
# ── Опции сборки ────────────────────────────────────────────────────────
option(BUILD_EXAMPLES "Сборка smoke-test программ из examples/" ON)
option(BUILD_TESTS "Сборка модульных и интеграционных тестов" OFF)
# ── Зависимости ─────────────────────────────────────────────────────────
# CUDA Toolkit — для cudart, заголовков cuda.h (Driver API), nvcc.
# NVENC SDK сам идёт через nv-codec-headers + dlopen libnvidia-encode.so,
# поэтому CUDA::nvenc намеренно НЕ линкуем (см. дизайн-документ часть 1.6).
find_package(CUDAToolkit 12.0 REQUIRED)
find_package(Threads REQUIRED)
# dl — для dlopen libnvidia-encode.so в runtime
find_library(LIBDL_LIBRARY dl REQUIRED)
# ── Сторонние библиотеки (subomodules в third_party/) ───────────────────
# cuframes — статически линкуем libcuframes. cuframes_static — это static lib
# который определён в third_party/cuframes/libcuframes/CMakeLists.txt.
# PIC обязателен — cuframes_static линкуется в наш SHARED libcuframes_composer.
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
set(BUILD_TESTING OFF CACHE BOOL "" FORCE)
set(BUILD_EXAMPLES OFF CACHE BOOL "Cuframes examples" FORCE)
set(BUILD_TOOLS OFF CACHE BOOL "Cuframes tools" FORCE)
set(BUILD_FFMPEG_FILTER OFF CACHE BOOL "" FORCE)
set(BUILD_PYTHON_BINDINGS OFF CACHE BOOL "" FORCE)
add_subdirectory(third_party/cuframes)
# Восстанавливаем BUILD_EXAMPLES для наших собственных examples/
set(BUILD_EXAMPLES ON CACHE BOOL "" FORCE)
# nv-codec-headers — header-only, нужны заголовки nvEncodeAPI.h
# (NVENC) и cuviddec.h (NVDEC, на будущее). Не как target, просто include path.
set(NVCODEC_HEADERS_DIR
"${CMAKE_CURRENT_SOURCE_DIR}/third_party/nv-codec-headers/include")
if(NOT EXISTS "${NVCODEC_HEADERS_DIR}/ffnvcodec/nvEncodeAPI.h")
message(FATAL_ERROR
"nv-codec-headers заголовки не найдены в ${NVCODEC_HEADERS_DIR}/ffnvcodec/. "
"Выполни: git submodule update --init --recursive")
endif()
# ── Подпроекты ──────────────────────────────────────────────────────────
# src/ — основная библиотека композитора
add_subdirectory(src)
# examples/ — smoke-test программы (по фазам разработки)
if(BUILD_EXAMPLES)
add_subdirectory(examples)
endif()
# ── Сводка конфигурации ─────────────────────────────────────────────────
message(STATUS "")
message(STATUS "cuframes-composer ${PROJECT_VERSION} конфигурация:")
message(STATUS " Build type: ${CMAKE_BUILD_TYPE}")
message(STATUS " CUDA Toolkit: ${CUDAToolkit_VERSION}")
message(STATUS " CUDA include: ${CUDAToolkit_INCLUDE_DIRS}")
message(STATUS " nv-codec-headers: ${NVCODEC_HEADERS_DIR}")
message(STATUS " BUILD_EXAMPLES: ${BUILD_EXAMPLES}")
message(STATUS " BUILD_TESTS: ${BUILD_TESTS}")
message(STATUS "")
+501
View File
@@ -0,0 +1,501 @@
GNU LESSER GENERAL PUBLIC LICENSE
Version 2.1, February 1999
Copyright (C) 1991, 1999 Free Software Foundation, Inc.
<https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
[This is the first released version of the Lesser GPL. It also counts
as the successor of the GNU Library Public License, version 2, hence
the version number 2.1.]
Preamble
The licenses for most software are designed to take away your
freedom to share and change it. By contrast, the GNU General Public
Licenses are intended to guarantee your freedom to share and change
free software--to make sure the software is free for all its users.
This license, the Lesser General Public License, applies to some
specially designated software packages--typically libraries--of the
Free Software Foundation and other authors who decide to use it. You
can use it too, but we suggest you first think carefully about whether
this license or the ordinary General Public License is the better
strategy to use in any particular case, based on the explanations below.
When we speak of free software, we are referring to freedom of use,
not price. Our General Public Licenses are designed to make sure that
you have the freedom to distribute copies of free software (and charge
for this service if you wish); that you receive source code or can get
it if you want it; that you can change the software and use pieces of
it in new free programs; and that you are informed that you can do
these things.
To protect your rights, we need to make restrictions that forbid
distributors to deny you these rights or to ask you to surrender these
rights. These restrictions translate to certain responsibilities for
you if you distribute copies of the library or if you modify it.
For example, if you distribute copies of the library, whether gratis
or for a fee, you must give the recipients all the rights that we gave
you. You must make sure that they, too, receive or can get the source
code. If you link other code with the library, you must provide
complete object files to the recipients, so that they can relink them
with the library after making changes to the library and recompiling
it. And you must show them these terms so they know their rights.
We protect your rights with a two-step method: (1) we copyright the
library, and (2) we offer you this license, which gives you legal
permission to copy, distribute and/or modify the library.
To protect each distributor, we want to make it very clear that
there is no warranty for the free library. Also, if the library is
modified by someone else and passed on, the recipients should know
that what they have is not the original version, so that the original
author's reputation will not be affected by problems that might be
introduced by others.
Finally, software patents pose a constant threat to the existence of
any free program. We wish to make sure that a company cannot
effectively restrict the users of a free program by obtaining a
restrictive license from a patent holder. Therefore, we insist that
any patent license obtained for a version of the library must be
consistent with the full freedom of use specified in this license.
Most GNU software, including some libraries, is covered by the
ordinary GNU General Public License. This license, the GNU Lesser
General Public License, applies to certain designated libraries, and
is quite different from the ordinary General Public License. We use
this license for certain libraries in order to permit linking those
libraries into non-free programs.
When a program is linked with a library, whether statically or using
a shared library, the combination of the two is legally speaking a
combined work, a derivative of the original library. The ordinary
General Public License therefore permits such linking only if the
entire combination fits its criteria of freedom. The Lesser General
Public License permits more lax criteria for linking other code with
the library.
We call this license the "Lesser" General Public License because it
does Less to protect the user's freedom than the ordinary General
Public License. It also provides other free software developers Less
of an advantage over competing non-free programs. These disadvantages
are the reason we use the ordinary General Public License for many
libraries. However, the Lesser license provides advantages in certain
special circumstances.
For example, on rare occasions, there may be a special need to
encourage the widest possible use of a certain library, so that it becomes
a de-facto standard. To achieve this, non-free programs must be
allowed to use the library. A more frequent case is that a free
library does the same job as widely used non-free libraries. In this
case, there is little to gain by limiting the free library to free
software only, so we use the Lesser General Public License.
In other cases, permission to use a particular library in non-free
programs enables a greater number of people to use a large body of
free software. For example, permission to use the GNU C Library in
non-free programs enables many more people to use the whole GNU
operating system, as well as its variant, the GNU/Linux operating
system.
Although the Lesser General Public License is Less protective of the
users' freedom, it does ensure that the user of a program that is
linked with the Library has the freedom and the wherewithal to run
that program using a modified version of the Library.
The precise terms and conditions for copying, distribution and
modification follow. Pay close attention to the difference between a
"work based on the library" and a "work that uses the library". The
former contains code derived from the library, whereas the latter must
be combined with the library in order to run.
GNU LESSER GENERAL PUBLIC LICENSE
TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
0. This License Agreement applies to any software library or other
program which contains a notice placed by the copyright holder or
other authorized party saying it may be distributed under the terms of
this Lesser General Public License (also called "this License").
Each licensee is addressed as "you".
A "library" means a collection of software functions and/or data
prepared so as to be conveniently linked with application programs
(which use some of those functions and data) to form executables.
The "Library", below, refers to any such software library or work
which has been distributed under these terms. A "work based on the
Library" means either the Library or any derivative work under
copyright law: that is to say, a work containing the Library or a
portion of it, either verbatim or with modifications and/or translated
straightforwardly into another language. (Hereinafter, translation is
included without limitation in the term "modification".)
"Source code" for a work means the preferred form of the work for
making modifications to it. For a library, complete source code means
all the source code for all modules it contains, plus any associated
interface definition files, plus the scripts used to control compilation
and installation of the library.
Activities other than copying, distribution and modification are not
covered by this License; they are outside its scope. The act of
running a program using the Library is not restricted, and output from
such a program is covered only if its contents constitute a work based
on the Library (independent of the use of the Library in a tool for
writing it). Whether that is true depends on what the Library does
and what the program that uses the Library does.
1. You may copy and distribute verbatim copies of the Library's
complete source code as you receive it, in any medium, provided that
you conspicuously and appropriately publish on each copy an
appropriate copyright notice and disclaimer of warranty; keep intact
all the notices that refer to this License and to the absence of any
warranty; and distribute a copy of this License along with the
Library.
You may charge a fee for the physical act of transferring a copy,
and you may at your option offer warranty protection in exchange for a
fee.
2. You may modify your copy or copies of the Library or any portion
of it, thus forming a work based on the Library, and copy and
distribute such modifications or work under the terms of Section 1
above, provided that you also meet all of these conditions:
a) The modified work must itself be a software library.
b) You must cause the files modified to carry prominent notices
stating that you changed the files and the date of any change.
c) You must cause the whole of the work to be licensed at no
charge to all third parties under the terms of this License.
d) If a facility in the modified Library refers to a function or a
table of data to be supplied by an application program that uses
the facility, other than as an argument passed when the facility
is invoked, then you must make a good faith effort to ensure that,
in the event an application does not supply such function or
table, the facility still operates, and performs whatever part of
its purpose remains meaningful.
(For example, a function in a library to compute square roots has
a purpose that is entirely well-defined independent of the
application. Therefore, Subsection 2d requires that any
application-supplied function or table used by this function must
be optional: if the application does not supply it, the square
root function must still compute square roots.)
These requirements apply to the modified work as a whole. If
identifiable sections of that work are not derived from the Library,
and can be reasonably considered independent and separate works in
themselves, then this License, and its terms, do not apply to those
sections when you distribute them as separate works. But when you
distribute the same sections as part of a whole which is a work based
on the Library, the distribution of the whole must be on the terms of
this License, whose permissions for other licensees extend to the
entire whole, and thus to each and every part regardless of who wrote
it.
Thus, it is not the intent of this section to claim rights or contest
your rights to work written entirely by you; rather, the intent is to
exercise the right to control the distribution of derivative or
collective works based on the Library.
In addition, mere aggregation of another work not based on the Library
with the Library (or with a work based on the Library) on a volume of
a storage or distribution medium does not bring the other work under
the scope of this License.
3. You may opt to apply the terms of the ordinary GNU General Public
License instead of this License to a given copy of the Library. To do
this, you must alter all the notices that refer to this License, so
that they refer to the ordinary GNU General Public License, version 2,
instead of to this License. (If a newer version than version 2 of the
ordinary GNU General Public License has appeared, then you can specify
that version instead if you wish.) Do not make any other change in
these notices.
Once this change is made in a given copy, it is irreversible for
that copy, so the ordinary GNU General Public License applies to all
subsequent copies and derivative works made from that copy.
This option is useful when you wish to copy part of the code of
the Library into a program that is not a library.
4. You may copy and distribute the Library (or a portion or
derivative of it, under Section 2) in object code or executable form
under the terms of Sections 1 and 2 above provided that you accompany
it with the complete corresponding machine-readable source code, which
must be distributed under the terms of Sections 1 and 2 above on a
medium customarily used for software interchange.
If distribution of object code is made by offering access to copy
from a designated place, then offering equivalent access to copy the
source code from the same place satisfies the requirement to
distribute the source code, even though third parties are not
compelled to copy the source along with the object code.
5. A program that contains no derivative of any portion of the
Library, but is designed to work with the Library by being compiled or
linked with it, is called a "work that uses the Library". Such a
work, in isolation, is not a derivative work of the Library, and
therefore falls outside the scope of this License.
However, linking a "work that uses the Library" with the Library
creates an executable that is a derivative of the Library (because it
contains portions of the Library), rather than a "work that uses the
library". The executable is therefore covered by this License.
Section 6 states terms for distribution of such executables.
When a "work that uses the Library" uses material from a header file
that is part of the Library, the object code for the work may be a
derivative work of the Library even though the source code is not.
Whether this is true is especially significant if the work can be
linked without the Library, or if the work is itself a library. The
threshold for this to be true is not precisely defined by law.
If such an object file uses only numerical parameters, data
structure layouts and accessors, and small macros and small inline
functions (ten lines or less in length), then the use of the object
file is unrestricted, regardless of whether it is legally a derivative
work. (Executables containing this object code plus portions of the
Library will still fall under Section 6.)
Otherwise, if the work is a derivative of the Library, you may
distribute the object code for the work under the terms of Section 6.
Any executables containing that work also fall under Section 6,
whether or not they are linked directly with the Library itself.
6. As an exception to the Sections above, you may also combine or
link a "work that uses the Library" with the Library to produce a
work containing portions of the Library, and distribute that work
under terms of your choice, provided that the terms permit
modification of the work for the customer's own use and reverse
engineering for debugging such modifications.
You must give prominent notice with each copy of the work that the
Library is used in it and that the Library and its use are covered by
this License. You must supply a copy of this License. If the work
during execution displays copyright notices, you must include the
copyright notice for the Library among them, as well as a reference
directing the user to the copy of this License. Also, you must do one
of these things:
a) Accompany the work with the complete corresponding
machine-readable source code for the Library including whatever
changes were used in the work (which must be distributed under
Sections 1 and 2 above); and, if the work is an executable linked
with the Library, with the complete machine-readable "work that
uses the Library", as object code and/or source code, so that the
user can modify the Library and then relink to produce a modified
executable containing the modified Library. (It is understood
that the user who changes the contents of definitions files in the
Library will not necessarily be able to recompile the application
to use the modified definitions.)
b) Use a suitable shared library mechanism for linking with the
Library. A suitable mechanism is one that (1) uses at run time a
copy of the library already present on the user's computer system,
rather than copying library functions into the executable, and (2)
will operate properly with a modified version of the library, if
the user installs one, as long as the modified version is
interface-compatible with the version that the work was made with.
c) Accompany the work with a written offer, valid for at
least three years, to give the same user the materials
specified in Subsection 6a, above, for a charge no more
than the cost of performing this distribution.
d) If distribution of the work is made by offering access to copy
from a designated place, offer equivalent access to copy the above
specified materials from the same place.
e) Verify that the user has already received a copy of these
materials or that you have already sent this user a copy.
For an executable, the required form of the "work that uses the
Library" must include any data and utility programs needed for
reproducing the executable from it. However, as a special exception,
the materials to be distributed need not include anything that is
normally distributed (in either source or binary form) with the major
components (compiler, kernel, and so on) of the operating system on
which the executable runs, unless that component itself accompanies
the executable.
It may happen that this requirement contradicts the license
restrictions of other proprietary libraries that do not normally
accompany the operating system. Such a contradiction means you cannot
use both them and the Library together in an executable that you
distribute.
7. You may place library facilities that are a work based on the
Library side-by-side in a single library together with other library
facilities not covered by this License, and distribute such a combined
library, provided that the separate distribution of the work based on
the Library and of the other library facilities is otherwise
permitted, and provided that you do these two things:
a) Accompany the combined library with a copy of the same work
based on the Library, uncombined with any other library
facilities. This must be distributed under the terms of the
Sections above.
b) Give prominent notice with the combined library of the fact
that part of it is a work based on the Library, and explaining
where to find the accompanying uncombined form of the same work.
8. You may not copy, modify, sublicense, link with, or distribute
the Library except as expressly provided under this License. Any
attempt otherwise to copy, modify, sublicense, link with, or
distribute the Library is void, and will automatically terminate your
rights under this License. However, parties who have received copies,
or rights, from you under this License will not have their licenses
terminated so long as such parties remain in full compliance.
9. You are not required to accept this License, since you have not
signed it. However, nothing else grants you permission to modify or
distribute the Library or its derivative works. These actions are
prohibited by law if you do not accept this License. Therefore, by
modifying or distributing the Library (or any work based on the
Library), you indicate your acceptance of this License to do so, and
all its terms and conditions for copying, distributing or modifying
the Library or works based on it.
10. Each time you redistribute the Library (or any work based on the
Library), the recipient automatically receives a license from the
original licensor to copy, distribute, link with or modify the Library
subject to these terms and conditions. You may not impose any further
restrictions on the recipients' exercise of the rights granted herein.
You are not responsible for enforcing compliance by third parties with
this License.
11. If, as a consequence of a court judgment or allegation of patent
infringement or for any other reason (not limited to patent issues),
conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot
distribute so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you
may not distribute the Library at all. For example, if a patent
license would not permit royalty-free redistribution of the Library by
all those who receive copies directly or indirectly through you, then
the only way you could satisfy both it and this License would be to
refrain entirely from distribution of the Library.
If any portion of this section is held invalid or unenforceable under any
particular circumstance, the balance of the section is intended to apply,
and the section as a whole is intended to apply in other circumstances.
It is not the purpose of this section to induce you to infringe any
patents or other property right claims or to contest validity of any
such claims; this section has the sole purpose of protecting the
integrity of the free software distribution system which is
implemented by public license practices. Many people have made
generous contributions to the wide range of software distributed
through that system in reliance on consistent application of that
system; it is up to the author/donor to decide if he or she is willing
to distribute software through any other system and a licensee cannot
impose that choice.
This section is intended to make thoroughly clear what is believed to
be a consequence of the rest of this License.
12. If the distribution and/or use of the Library is restricted in
certain countries either by patents or by copyrighted interfaces, the
original copyright holder who places the Library under this License may add
an explicit geographical distribution limitation excluding those countries,
so that distribution is permitted only in or among countries not thus
excluded. In such case, this License incorporates the limitation as if
written in the body of this License.
13. The Free Software Foundation may publish revised and/or new
versions of the Lesser General Public License from time to time.
Such new versions will be similar in spirit to the present version,
but may differ in detail to address new problems or concerns.
Each version is given a distinguishing version number. If the Library
specifies a version number of this License which applies to it and
"any later version", you have the option of following the terms and
conditions either of that version or of any later version published by
the Free Software Foundation. If the Library does not specify a
license version number, you may choose any version ever published by
the Free Software Foundation.
14. If you wish to incorporate parts of the Library into other free
programs whose distribution conditions are incompatible with these,
write to the author to ask for permission. For software which is
copyrighted by the Free Software Foundation, write to the Free
Software Foundation; we sometimes make exceptions for this. Our
decision will be guided by the two goals of preserving the free status
of all derivatives of our free software and of promoting the sharing
and reuse of software generally.
NO WARRANTY
15. BECAUSE THE LIBRARY IS LICENSED FREE OF CHARGE, THERE IS NO
WARRANTY FOR THE LIBRARY, TO THE EXTENT PERMITTED BY APPLICABLE LAW.
EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR
OTHER PARTIES PROVIDE THE LIBRARY "AS IS" WITHOUT WARRANTY OF ANY
KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE
LIBRARY IS WITH YOU. SHOULD THE LIBRARY PROVE DEFECTIVE, YOU ASSUME
THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN
WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY
AND/OR REDISTRIBUTE THE LIBRARY AS PERMITTED ABOVE, BE LIABLE TO YOU
FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR
CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE
LIBRARY (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING
RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A
FAILURE OF THE LIBRARY TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF
SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
DAMAGES.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Libraries
If you develop a new library, and you want it to be of the greatest
possible use to the public, we recommend making it free software that
everyone can redistribute and change. You can do so by permitting
redistribution under these terms (or, alternatively, under the terms of the
ordinary General Public License).
To apply these terms, attach the following notices to the library. It is
safest to attach them to the start of each source file to most effectively
convey the exclusion of warranty; and each file should have at least the
"copyright" line and a pointer to where the full notice is found.
<one line to give the library's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 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
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
You should also get your employer (if you work as a programmer) or your
school, if any, to sign a "copyright disclaimer" for the library, if
necessary. Here is a sample; alter the names:
Yoyodyne, Inc., hereby disclaims all copyright interest in the
library `Frob' (a library for tweaking knobs) written by James Random Hacker.
<signature of Moe Ghoul>, 1 April 1990
Moe Ghoul, President of Vice
That's all there is to it!
+55
View File
@@ -0,0 +1,55 @@
# cuframes-composer
Стандалонный композитор-демон для multi-source видео grid через CUDA + NVENC + RTSP.
Заменяет монолитный ffmpeg-конвейер (`ffmpeg + vf_cuda_grid` фильтр) для случаев, когда нужно:
- Поток продолжает работать при потере любого числа источников (graceful degradation)
- Композитор сам управляет частотой кадров и обработкой ошибок без зависимости от семантики ffmpeg-демухера
- Минимум перемещений данных: zero-copy CUDA от источника `cuframes` напрямую в NVENC
## Статус
**Phase 1 — MVP.** В разработке. Не для боевой эксплуатации.
См. [дизайн-документ](https://git.goldix.org/gx/cuframes/raw/branch/main/docs/DESIGN-composer-daemon.md) для архитектурных решений и поэтапного плана.
## Зависимости
- [cuframes](https://git.goldix.org/gx/cuframes) — библиотека zero-copy передачи кадров. Подключена как git submodule.
- [nv-codec-headers](https://github.com/FFmpeg/nv-codec-headers) — MIT-licensed заголовки NVENC API. Подключена как git submodule. Сама библиотека `libnvidia-encode.so` грузится через `dlopen` при старте (это даёт LGPL-совместимость — см. дизайн-документ часть 1.6).
- CUDA Toolkit 12.x+ (для cuda runtime и компиляции)
- NVIDIA драйвер 525+ (для NVENC и `cuMemCreate` POSIX FD)
- Linux 64-bit (POSIX shm, SCM_RIGHTS)
Дополнительно по фазам:
- Phase 3: `libfreetype` (текст), `lodepng` через submodule (PNG-декодирование)
- Phase 4: `libzmq` (управление)
## Сборка
```bash
git clone --recursive git@git.goldix.org:gx/cuframes-composer.git
cd cuframes-composer
cmake -B build -G Ninja
ninja -C build
```
## Поэтапный план
| фаза | срок | результат |
|---|---|---|
| 1 | 1 неделя | один источник → NVENC → файл .h264 (доказательство zero-copy) |
| 2 | 2 недели | четыре источника + композиция через `libcugrid` |
| 3 | 2 недели | оверлеи + RTSP push к mediamtx + AAC passthrough из `/live-audio` |
| 4 | 1 неделя | паритет ZMQ-управления с фильтром `vf_cuda_grid` |
| 5 | 1 неделя | боевое развёртывание + MQTT health + watchdog |
| 6 | 2 недели | тесты + бенчмарки + документация |
Итого ~9 недель для одного разработчика.
## Лицензия
LGPL-2.1-or-later. См. файл [LICENSE](LICENSE).
NVENC SDK headers (`third_party/nv-codec-headers`) — MIT license, совместима с LGPL.
+12
View File
@@ -0,0 +1,12 @@
# Smoke-test программы — по одной на каждую фазу разработки.
#
# Phase 1: simple_record — подключиться к одному cuframes публишеру,
# закодировать через NVENC, записать H.264 в файл. Доказывает что
# zero-copy NVENC из VMM-памяти cuframes работает.
#
# Phase 2: composer_4_to_file — 4 источника + композиция → файл.
# Phase 3: composer_rtsp — 4 источника + композиция → RTSP push.
add_executable(simple_record simple_record.c)
target_link_libraries(simple_record PRIVATE cuframes_composer_static)
target_include_directories(simple_record PRIVATE ${CMAKE_SOURCE_DIR}/include)
+270
View File
@@ -0,0 +1,270 @@
/* simple_record — Phase 1 smoke test.
*
* Подписывается на один cuframes-источник, кодирует каждый кадр через
* NVENC и пишет H.264 Annex-B byte stream в файл. Завершение по SIGINT.
*
* Цель — доказать end-to-end zero-copy: NV12 frame из VMM-памяти publisher'а
* напрямую попадает в NVENC без промежуточных копий через CPU.
*
* Использование:
* simple_record --key cam-parking --out test.h264 [--fps 25] [--bitrate 5000]
*
* Полученный файл проверяется через ffmpeg:
* ffprobe test.h264 → должно показать h264 stream
* ffmpeg -i test.h264 ... → должен декодироваться
*
* Лицензия: LGPL-2.1+
*/
#include "../include/cuframes_composer/nvenc.h"
#include "../include/cuframes_composer/source.h"
#include <cuda.h>
#include <errno.h>
#include <getopt.h>
#include <signal.h>
#include <stdatomic.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
static volatile sig_atomic_t g_stop = 0;
static void on_sigint(int sig)
{
(void)sig;
g_stop = 1;
}
/* user-data, передаваемая encoder callback'у */
typedef struct write_ctx {
FILE *fp;
uint64_t bytes_written;
uint64_t frames_encoded;
uint64_t idr_count;
} write_ctx_t;
/* Encoder callback — пишем H.264 Annex-B bytes как есть. */
static void on_bitstream(const uint8_t *bs, size_t size, int64_t pts_ns,
int is_idr, void *user)
{
(void)pts_ns;
write_ctx_t *ctx = (write_ctx_t *)user;
if (fwrite(bs, 1, size, ctx->fp) != size) {
fprintf(stderr, "[simple_record] fwrite failed: %s\n", strerror(errno));
} else {
ctx->bytes_written += size;
ctx->frames_encoded++;
if (is_idr) ctx->idr_count++;
}
}
static const char *cu_err(CUresult r)
{
const char *s = NULL;
cuGetErrorString(r, &s);
return s ? s : "unknown";
}
#define CUCHECK(expr) do { \
CUresult _r = (expr); \
if (_r != CUDA_SUCCESS) { \
fprintf(stderr, "[simple_record] %s failed: %s\n", #expr, cu_err(_r)); \
return 1; \
} \
} while (0)
int main(int argc, char **argv)
{
const char *key = NULL;
const char *out_path = NULL;
int fps = 25;
int bitrate_kbps = 5000;
int max_seconds = 0; /* 0 = до SIGINT */
static struct option opts[] = {
{"key", required_argument, 0, 'k'},
{"out", required_argument, 0, 'o'},
{"fps", required_argument, 0, 'f'},
{"bitrate", required_argument, 0, 'b'},
{"seconds", required_argument, 0, 's'},
{"help", no_argument, 0, 'h'},
{0, 0, 0, 0},
};
int c;
while ((c = getopt_long(argc, argv, "k:o:f:b:s:h", opts, NULL)) != -1) {
switch (c) {
case 'k': key = optarg; break;
case 'o': out_path = optarg; break;
case 'f': fps = atoi(optarg); break;
case 'b': bitrate_kbps = atoi(optarg); break;
case 's': max_seconds = atoi(optarg); break;
case 'h':
default:
fprintf(stderr,
"Использование: %s --key <cuframes-key> --out <file.h264>\n"
" [--fps 25] [--bitrate 5000] [--seconds N]\n",
argv[0]);
return c == 'h' ? 0 : 1;
}
}
if (!key || !out_path) {
fprintf(stderr, "[simple_record] требуются --key и --out\n");
return 1;
}
signal(SIGINT, on_sigint);
signal(SIGTERM, on_sigint);
/* 1) CUDA primary context на устройстве 0. cuframes-subscriber и NVENC
* оба должны работать в одном контексте. */
CUCHECK(cuInit(0));
CUdevice dev;
CUCHECK(cuDeviceGet(&dev, 0));
CUcontext ctx;
CUCHECK(cuDevicePrimaryCtxRetain(&ctx, dev));
CUCHECK(cuCtxPushCurrent(ctx));
/* 2) Открыть source. */
cfc_source_config_t scfg = {
.key = key,
.consumer_name = "simple_record",
.cuda_device = 0,
};
cfc_source_t *src = NULL;
if (cfc_source_create(&scfg, &src) != 0) {
fprintf(stderr, "[simple_record] cfc_source_create failed\n");
return 1;
}
/* 3) Ждать первый кадр чтобы узнать размер. До 30 секунд. */
cfc_source_snapshot_t snap = { 0 };
int waited_ms = 0;
while (!g_stop && waited_ms < 30000) {
cfc_source_get_latest(src, &snap);
if (snap.state == CFC_SOURCE_ACTIVE && snap.width > 0) break;
struct timespec ts = {.tv_sec = 0, .tv_nsec = 50 * 1000 * 1000};
nanosleep(&ts, NULL);
waited_ms += 50;
}
if (g_stop) goto cleanup_src;
if (snap.width <= 0 || snap.height <= 0) {
fprintf(stderr,
"[simple_record] не дождался первого кадра за 30с (state=%d)\n",
snap.state);
goto cleanup_src;
}
fprintf(stderr,
"[simple_record] первый кадр: %dx%d pitch=%d → создаю encoder\n",
snap.width, snap.height, snap.pitch_y);
/* 4) Создать encoder под полученный размер. */
cfc_encoder_config_t ecfg = {
.cuda_ctx = ctx,
.width = snap.width,
.height = snap.height,
.fps_num = fps,
.fps_den = 1,
.bitrate_kbps = bitrate_kbps,
.gop_size = fps, /* IDR раз в секунду — стандарт для RTSP */
.num_b_frames = 0, /* low-latency: B-кадры мешают */
.preset = "ll",
};
cfc_encoder_t *enc = NULL;
if (cfc_encoder_create(&ecfg, &enc) != 0) {
fprintf(stderr, "[simple_record] cfc_encoder_create failed\n");
goto cleanup_src;
}
/* 5) Открыть выходной файл. */
write_ctx_t wctx = { 0 };
wctx.fp = fopen(out_path, "wb");
if (!wctx.fp) {
fprintf(stderr, "[simple_record] fopen(%s) failed: %s\n",
out_path, strerror(errno));
goto cleanup_enc;
}
/* 6) Главный цикл — забираем кадры по seq, кодируем. */
uint64_t last_seq = 0;
int64_t start_us;
struct timespec ts_start;
clock_gettime(CLOCK_MONOTONIC, &ts_start);
start_us = (int64_t)ts_start.tv_sec * 1000000 + ts_start.tv_nsec / 1000;
fprintf(stderr, "[simple_record] начало записи в %s (Ctrl+C для остановки)\n",
out_path);
while (!g_stop) {
cfc_source_get_latest(src, &snap);
if (snap.state != CFC_SOURCE_ACTIVE) {
/* Источник не active — короткий sleep и снова. */
struct timespec ts = {.tv_sec = 0, .tv_nsec = 10 * 1000 * 1000};
nanosleep(&ts, NULL);
continue;
}
if (snap.seq == last_seq) {
/* Тот же кадр что и был — ждём новый. Спим четверть кадрового
* интервала чтобы не крутить CPU впустую. */
int sleep_ns = 1000000000 / fps / 4;
struct timespec ts = {.tv_sec = 0, .tv_nsec = sleep_ns};
nanosleep(&ts, NULL);
continue;
}
last_seq = snap.seq;
if (cfc_encoder_encode_frame(enc, snap.ptr, snap.pitch_y,
snap.pts_ns, on_bitstream, &wctx) != 0) {
fprintf(stderr, "[simple_record] encode_frame failed\n");
break;
}
/* Прогресс каждые 100 кадров. */
if (wctx.frames_encoded > 0 && wctx.frames_encoded % 100 == 0) {
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
int64_t now_us = (int64_t)now.tv_sec * 1000000 + now.tv_nsec / 1000;
double elapsed_s = (now_us - start_us) / 1e6;
fprintf(stderr,
"[simple_record] %llu кадров, %llu IDR, %.1f МБ за %.1fс (%.1f fps)\n",
(unsigned long long)wctx.frames_encoded,
(unsigned long long)wctx.idr_count,
wctx.bytes_written / 1048576.0,
elapsed_s,
wctx.frames_encoded / elapsed_s);
}
if (max_seconds > 0) {
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
int64_t elapsed = ((int64_t)now.tv_sec * 1000000 + now.tv_nsec / 1000) - start_us;
if (elapsed / 1000000 >= max_seconds) {
fprintf(stderr, "[simple_record] достигнут лимит %dс\n", max_seconds);
break;
}
}
}
fprintf(stderr, "[simple_record] flush encoder\n");
cfc_encoder_flush(enc, on_bitstream, &wctx);
fprintf(stderr,
"[simple_record] итого: %llu кадров, %llu IDR, %.2f МБ\n",
(unsigned long long)wctx.frames_encoded,
(unsigned long long)wctx.idr_count,
wctx.bytes_written / 1048576.0);
fclose(wctx.fp);
cleanup_enc:
cfc_encoder_destroy(enc);
cleanup_src:
cfc_source_destroy(src);
cuCtxPopCurrent(NULL);
cuDevicePrimaryCtxRelease(dev);
return 0;
}
+107
View File
@@ -0,0 +1,107 @@
/* cuframes-composer — обвязка вокруг NVIDIA NVENC API.
*
* Динамически грузит libnvidia-encode.so через dlopen, чтобы пакет
* cuframes-composer оставался под LGPL-2.1+ без статической линковки
* проприетарного SDK. См. дизайн-документ часть 1.6.
*
* Принимает на вход CUdeviceptr на NV12 frame (zero-copy через
* NV_ENC_INPUT_RESOURCE_TYPE_CUDADEVICEPTR + nvEncRegisterResource).
* Выдаёт сжатый H.264 bitstream через callback (caller записывает его
* в файл / RTP-пакетизирует / etc).
*
* Lifecycle:
* create(cfg) — open session, init encoder
* encode_frame(...) — на каждый входной кадр (один CUdeviceptr)
* flush(...) — в конце потока, чтобы вытащить остатки B-кадров
* destroy(...) — закрыть session, выгрузить SDK
*
* Поток должен быть single-threaded для одного encoder'а (NVENC API
* не реентрабельный для одной сессии).
*
* Лицензия: LGPL-2.1+
*/
#ifndef CUFRAMES_COMPOSER_NVENC_H
#define CUFRAMES_COMPOSER_NVENC_H
#include <cuda.h>
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
/* Параметры кодировщика. Для Phase 1 минимальный набор; в будущих фазах
* будут расширяться для RTSP (rate control, GOP, intra-refresh, и т.п.). */
typedef struct cfc_encoder_config {
CUcontext cuda_ctx; /* CUDA-контекст, в котором лежат входные VMM-буферы */
int32_t width; /* ширина кадра в пикселях */
int32_t height; /* высота кадра в пикселях */
int32_t fps_num; /* числитель частоты кадров (25) */
int32_t fps_den; /* знаменатель частоты кадров (1) */
int32_t bitrate_kbps; /* битрейт в килобитах в секунду (5000 = 5 Мбит/с) */
int32_t gop_size; /* интервал между keyframe'ами в кадрах (25 = 1 IDR в секунду) */
int32_t num_b_frames; /* B-кадры (0 для low-latency RTSP) */
/* Пресет — соответствует NV_ENC_TUNING_INFO + preset GUID.
* "ll" = low-latency, "p4" = preset P4 (баланс), "p7" = highest quality. */
const char *preset; /* "ll", "p4", "p7" — по умолчанию "ll" */
} cfc_encoder_config_t;
typedef struct cfc_encoder cfc_encoder_t;
/* Callback для записанного H.264 bitstream'а. Вызывается синхронно из
* cfc_encoder_encode_frame / cfc_encoder_flush. Указатель действителен только
* на время вызова (буфер NVENC будет разблокирован после возврата). */
typedef void (*cfc_encoder_output_cb)(
const uint8_t *bitstream, /* данные H.264 (Annex B byte stream) */
size_t size, /* размер в байтах */
int64_t pts_ns, /* presentation timestamp (передан в encode_frame) */
int is_idr, /* 1 если кадр IDR (keyframe) */
void *user /* user data, переданный в encode_frame */
);
/* Создать encoder. Возвращает 0 при успехе. */
int cfc_encoder_create(const cfc_encoder_config_t *cfg, cfc_encoder_t **out);
/* Закодировать один кадр. ptr — CUdeviceptr на начало NV12 frame в VRAM,
* pitch — ширина строки в байтах (для NV12 равна width).
*
* Callback может быть вызван 0 или 1 раз для одного encode_frame: NVENC
* может буферизовать кадры внутри (особенно при B-кадрах). Чтобы вытащить
* последние буферизованные — вызвать flush в конце потока. */
int cfc_encoder_encode_frame(
cfc_encoder_t *enc,
CUdeviceptr ptr,
int pitch,
int64_t pts_ns,
cfc_encoder_output_cb cb,
void *user
);
/* Завершить поток. Передаёт NVENC флаг end-of-stream, выдаёт оставшиеся
* закодированные кадры через callback. После flush encoder можно либо
* destroy, либо использовать снова с новым GOP'ом. */
int cfc_encoder_flush(
cfc_encoder_t *enc,
cfc_encoder_output_cb cb,
void *user
);
/* Выдать SPS/PPS — заголовки H.264 sequence/picture parameter sets.
* Нужны для записи в начало MP4-контейнера или для отправки в SDP при RTSP. */
int cfc_encoder_get_sequence_params(
cfc_encoder_t *enc,
uint8_t *out, /* буфер caller'а */
size_t *inout_size /* при вызове — размер буфера, при возврате — реальный размер */
);
/* Закрыть encoder, выгрузить SDK. */
int cfc_encoder_destroy(cfc_encoder_t *enc);
#ifdef __cplusplus
}
#endif
#endif /* CUFRAMES_COMPOSER_NVENC_H */
+84
View File
@@ -0,0 +1,84 @@
/* cuframes-composer — обвязка над cuframes_subscriber с явной машиной
* состояний и автоматическим переподключением. Создаёт собственный поток
* для блокирующего cuframes_subscriber_next и держит снимок последнего
* успешного кадра, который читатели берут не блокируясь.
*
* Используется в композиторе для каждого входного источника (одна камера).
*
* Lifecycle:
* create() → запускается поток, который начинает с DISCONNECTED, идёт
* в CONNECTING (попытка подписаться), затем ACTIVE при успехе
* get_latest() → не блокируется, возвращает снимок последнего кадра
* + текущее состояние
* destroy() → останавливает поток, освобождает ресурсы cuframes
*
* Thread safety: create/destroy — main thread. get_latest — любой поток.
*
* Лицензия: LGPL-2.1+
*/
#ifndef CUFRAMES_COMPOSER_SOURCE_H
#define CUFRAMES_COMPOSER_SOURCE_H
#include <cuda.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
/* Состояние источника. ACTIVE — единственное «здоровое» состояние,
* остальные означают какую-то форму потери связи или ожидания. */
typedef enum cfc_source_state {
CFC_SOURCE_DISCONNECTED = 0, /* стартовое или после destroy подписки */
CFC_SOURCE_CONNECTING, /* идёт subscribe handshake */
CFC_SOURCE_ACTIVE, /* получаем кадры, последний < N мс назад */
CFC_SOURCE_STALE, /* подписка жива, но кадры не приходят */
CFC_SOURCE_DEAD, /* подписка отвалилась, ждём backoff до retry */
} cfc_source_state_t;
/* Конфигурация одного источника. */
typedef struct cfc_source_config {
const char *key; /* cuframes key (например "cam-parking") */
const char *consumer_name; /* имя для cuframes, должно быть уникально на publisher */
int32_t cuda_device; /* индекс CUDA-устройства, обычно 0 */
int32_t reconnect_min_ms; /* минимальный backoff при DEAD → CONNECTING (по умолчанию 1000) */
int32_t reconnect_max_ms; /* максимальный backoff (по умолчанию 30000) */
int32_t stale_threshold_ms; /* без кадра > N → ACTIVE → STALE (по умолчанию 500) */
int32_t dead_threshold_ms; /* без кадра > N → STALE → DEAD (по умолчанию 5000) */
} cfc_source_config_t;
typedef struct cfc_source cfc_source_t;
/* Снимок последнего успешного кадра. ptr указывает на VMM-mapped CUDA-память
* cuframes publisher'а. Он действителен до следующего вызова get_latest:
* после этого источник может перейти на следующий слот ring buffer'а. */
typedef struct cfc_source_snapshot {
CUdeviceptr ptr; /* указатель на NV12 frame (Y plane) */
int32_t width;
int32_t height;
int32_t pitch_y;
int32_t pitch_uv;
int64_t pts_ns; /* timestamp от publisher'а (CLOCK_MONOTONIC ns) */
uint64_t seq; /* sequence number */
cfc_source_state_t state; /* текущее состояние источника */
int64_t last_frame_age_us; /* сколько микросекунд назад последний успешный кадр; -1 если никогда */
} cfc_source_snapshot_t;
/* Создать источник. Возвращает 0 при успехе. Старт асинхронный:
* cfc_source_create возвращает сразу, фоновый поток выполняет subscribe.
* До первого успешного кадра состояние будет CONNECTING либо DEAD. */
int cfc_source_create(const cfc_source_config_t *cfg, cfc_source_t **out);
/* Получить снимок последнего кадра. Не блокируется. Возвращает 0 при успехе.
* Поле state в out выставляется всегда (даже если кадров ещё не было). */
int cfc_source_get_latest(cfc_source_t *src, cfc_source_snapshot_t *out);
/* Освободить источник. Останавливает поток (joins), отключает подписку. */
int cfc_source_destroy(cfc_source_t *src);
#ifdef __cplusplus
}
#endif
#endif /* CUFRAMES_COMPOSER_SOURCE_H */
+63
View File
@@ -0,0 +1,63 @@
# Основная библиотека композитора — `libcuframes_composer.so`. Содержит:
# - source.c подписка к cuframes публишеру + state machine
# - nvenc_loader.c dlopen libnvidia-encode.so + загрузка API таблицы
# - nvenc.c обвязка вокруг NVENC SDK (init/encode/teardown)
#
# Дальше по фазам:
# Phase 2: compose.c (CUDA composition), ringbuf.c (SPSC swap)
# Phase 3: rtsp_publisher.c, rtp_h264.c, overlay.c, png_decode.c, text_render.c
# Phase 4: control_zmq.c
# Phase 5: health_mqtt.c
set(COMPOSER_SOURCES
source.c
nvenc_loader.c
nvenc.c
)
add_library(cuframes_composer SHARED ${COMPOSER_SOURCES})
add_library(cuframes_composer_static STATIC ${COMPOSER_SOURCES})
foreach(target cuframes_composer cuframes_composer_static)
target_include_directories(${target}
PUBLIC
$<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/include>
$<INSTALL_INTERFACE:include>
PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}
${NVCODEC_HEADERS_DIR}
)
target_compile_features(${target} PRIVATE c_std_11)
target_compile_options(${target} PRIVATE
-Wall -Wextra -Wpedantic
$<$<CONFIG:Debug>:-O0 -g3>
$<$<CONFIG:Release>:-O2 -g>
)
target_link_libraries(${target}
PUBLIC
cuframes_static # из third_party/cuframes
CUDA::cudart
CUDA::cuda_driver
Threads::Threads
${LIBDL_LIBRARY} # для dlopen libnvidia-encode.so
rt
)
endforeach()
set_target_properties(cuframes_composer PROPERTIES
VERSION ${PROJECT_VERSION}
SOVERSION 0
)
# Install rules
include(GNUInstallDirs)
install(TARGETS cuframes_composer cuframes_composer_static
EXPORT cuframesComposerTargets
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
)
install(DIRECTORY ${CMAKE_SOURCE_DIR}/include/cuframes_composer
DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
FILES_MATCHING PATTERN "*.h"
)
+413
View File
@@ -0,0 +1,413 @@
/* NVENC обвязка — реализация публичного API cuframes_composer/nvenc.h.
*
* Lifecycle NVENC:
* 1) nvEncOpenEncodeSessionEx(CUDA context) → handle
* 2) nvEncGetEncodeGUIDs → проверить что H264 поддерживается
* 3) nvEncGetEncodePresetConfigEx(preset) → дефолтный config от пресета
* 4) Override полей config'а (bitrate, GOP, B-frames)
* 5) nvEncInitializeEncoder(params{config, GUID, размер, fps, ...})
* 6) Создать pool of output bitstream buffers (через CreateBitstreamBuffer)
*
* Encode loop:
* 1) nvEncRegisterResource(CUdeviceptr, pitch) → registered_resource
* (кешируем — один cuframes-slot регистрируется один раз и переиспользуется)
* 2) nvEncMapInputResource(registered) → NV_ENC_INPUT_PTR (mapped)
* 3) nvEncEncodePicture(picParams{mapped, output_buffer, pts, ...})
* может вернуть NV_ENC_ERR_NEED_MORE_INPUT — это OK, означает буферизация
* 4) Если encode успешен (NV_ENC_SUCCESS):
* nvEncLockBitstream → выгрузить bytes → callback → nvEncUnlockBitstream
* 5) nvEncUnmapInputResource
*
* Flush (end-of-stream):
* 1) encodePicture с completionEvent=NULL и pictureStruct=NV_ENC_PIC_STRUCT_FRAME
* + endOfStream=1 → вытащит буферизованные кадры
* 2) Lock/Unlock в цикле пока есть выходные кадры
*
* Phase 1 ограничения (упрощения, будут сняты в следующих фазах):
* - Один output bitstream buffer (нет async encode, sync flow)
* - Линейный кеш registered resources (для Phase 2 → hash map)
* - Без SEI / metadata insertion (для RTSP добавится в Phase 3)
* - Без adaptive bitrate (Phase 3)
*
* Reference: NVIDIA Video Codec SDK Sample код + ffmpeg/libavcodec/nvenc.c
* (использован как образец сборки picParams и initParams).
*/
#include "../include/cuframes_composer/nvenc.h"
#include "nvenc_loader.h"
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
/* Максимальный размер кеша зарегистрированных ресурсов. cuframes ring buffer
* для одного источника — 16 slots, для 4 источников — 64. Phase 1 = 16. */
#define CFC_MAX_REGISTERED_RESOURCES 64
/* Размер выходного bitstream buffer'а. Default NVENC = 1 MB / frame —
* хватает с запасом для 1080p H.264 даже на keyframe. */
#define CFC_BITSTREAM_BUFFER_SIZE (2 * 1024 * 1024)
typedef struct registered_resource {
CUdeviceptr ptr; /* ключ кеша */
int pitch;
NV_ENC_REGISTERED_PTR regptr; /* возвращается nvEncRegisterResource */
} registered_resource_t;
struct cfc_encoder {
void *session; /* NVENC session handle */
NV_ENC_BUFFER_FORMAT input_format; /* NV12 */
int width, height;
int fps_num, fps_den;
/* Output bitstream buffer (один на Phase 1, потом pool) */
NV_ENC_OUTPUT_PTR output_bitstream;
/* Кеш зарегистрированных входных буферов */
registered_resource_t registered[CFC_MAX_REGISTERED_RESOURCES];
int registered_count;
pthread_mutex_t registered_mu;
};
/* ── Helpers ──────────────────────────────────────────────────────────── */
/* Безопасно вернуть управление с ошибкой + лог. */
#define NVE_CHECK(expr, label) do { \
NVENCSTATUS _s = (expr); \
if (_s != NV_ENC_SUCCESS) { \
fprintf(stderr, "[cfc/nvenc] %s:%d %s → %s\n", \
__FILE__, __LINE__, #expr, cfc_nvenc_status_str(_s)); \
rc = -1; goto label; \
} \
} while (0)
/* GUID H.264 кодека. Это стандартная константа из NVENC SDK. */
static const GUID nv_codec_h264 = NV_ENC_CODEC_H264_GUID;
/* GUID пресетов — выбираются по строковому имени из cfc_encoder_config.preset. */
static const GUID nv_preset_p1 = NV_ENC_PRESET_P1_GUID; /* fastest */
static const GUID nv_preset_p4 = NV_ENC_PRESET_P4_GUID; /* balanced */
static const GUID nv_preset_p7 = NV_ENC_PRESET_P7_GUID; /* highest quality */
/* Найти/выбрать preset GUID по строковому имени. */
static const GUID *select_preset(const char *name)
{
if (!name || !*name || !strcmp(name, "ll") || !strcmp(name, "p4")) return &nv_preset_p4;
if (!strcmp(name, "p1")) return &nv_preset_p1;
if (!strcmp(name, "p7")) return &nv_preset_p7;
fprintf(stderr, "[cfc/nvenc] unknown preset '%s', using p4\n", name);
return &nv_preset_p4;
}
/* Регистрация CUDA-указателя в NVENC (или поиск в кеше). Возвращает
* NV_ENC_REGISTERED_PTR или NULL при ошибке. Thread-safe через mutex. */
static NV_ENC_REGISTERED_PTR get_or_register(cfc_encoder_t *enc,
CUdeviceptr ptr, int pitch)
{
pthread_mutex_lock(&enc->registered_mu);
/* Поиск в кеше */
for (int i = 0; i < enc->registered_count; i++) {
if (enc->registered[i].ptr == ptr && enc->registered[i].pitch == pitch) {
NV_ENC_REGISTERED_PTR rp = enc->registered[i].regptr;
pthread_mutex_unlock(&enc->registered_mu);
return rp;
}
}
if (enc->registered_count >= CFC_MAX_REGISTERED_RESOURCES) {
pthread_mutex_unlock(&enc->registered_mu);
fprintf(stderr, "[cfc/nvenc] registered cache overflow (max %d)\n",
CFC_MAX_REGISTERED_RESOURCES);
return NULL;
}
/* Регистрация. NV12 формат: Y plane на полном разрешении, UV plane
* на половинном (interleaved). Размер NV_ENC_REGISTER_RESOURCE — это
* вся NV12-плоскость (Y + UV, height * pitch * 1.5). */
NV_ENC_REGISTER_RESOURCE reg = { 0 };
reg.version = NV_ENC_REGISTER_RESOURCE_VER;
reg.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_CUDADEVICEPTR;
reg.width = enc->width;
reg.height = enc->height;
reg.pitch = pitch;
reg.resourceToRegister = (void *)ptr;
reg.bufferFormat = enc->input_format;
reg.bufferUsage = NV_ENC_INPUT_IMAGE;
NVENCSTATUS s = g_nvenc_funcs.nvEncRegisterResource(enc->session, &reg);
if (s != NV_ENC_SUCCESS) {
pthread_mutex_unlock(&enc->registered_mu);
fprintf(stderr,
"[cfc/nvenc] nvEncRegisterResource failed: %s (ptr=%p pitch=%d)\n",
cfc_nvenc_status_str(s), (void *)ptr, pitch);
return NULL;
}
int idx = enc->registered_count++;
enc->registered[idx].ptr = ptr;
enc->registered[idx].pitch = pitch;
enc->registered[idx].regptr = reg.registeredResource;
pthread_mutex_unlock(&enc->registered_mu);
return reg.registeredResource;
}
/* Вытащить готовый закодированный кадр и отдать через callback.
* Возвращает 0 если ничего нет (NEED_MORE_INPUT), 1 если выдал, -1 ошибка. */
static int drain_output(cfc_encoder_t *enc,
cfc_encoder_output_cb cb, void *user)
{
NV_ENC_LOCK_BITSTREAM lock = { 0 };
lock.version = NV_ENC_LOCK_BITSTREAM_VER;
lock.outputBitstream = enc->output_bitstream;
NVENCSTATUS s = g_nvenc_funcs.nvEncLockBitstream(enc->session, &lock);
if (s == NV_ENC_ERR_NEED_MORE_INPUT) {
return 0;
}
if (s != NV_ENC_SUCCESS) {
fprintf(stderr, "[cfc/nvenc] nvEncLockBitstream failed: %s\n",
cfc_nvenc_status_str(s));
return -1;
}
int is_idr = (lock.pictureType == NV_ENC_PIC_TYPE_IDR);
int64_t pts_ns = (int64_t)lock.outputTimeStamp;
if (cb) {
cb((const uint8_t *)lock.bitstreamBufferPtr, lock.bitstreamSizeInBytes,
pts_ns, is_idr, user);
}
s = g_nvenc_funcs.nvEncUnlockBitstream(enc->session, enc->output_bitstream);
if (s != NV_ENC_SUCCESS) {
fprintf(stderr, "[cfc/nvenc] nvEncUnlockBitstream failed: %s\n",
cfc_nvenc_status_str(s));
return -1;
}
return 1;
}
/* ── Public API ───────────────────────────────────────────────────────── */
int cfc_encoder_create(const cfc_encoder_config_t *cfg, cfc_encoder_t **out)
{
if (!cfg || !out) return -1;
if (cfg->width <= 0 || cfg->height <= 0) return -1;
if (cfg->fps_num <= 0 || cfg->fps_den <= 0) return -1;
if (cfg->bitrate_kbps <= 0) return -1;
if (cfc_nvenc_loader_init() != 0) return -1;
cfc_encoder_t *enc = calloc(1, sizeof(*enc));
if (!enc) return -1;
enc->width = cfg->width;
enc->height = cfg->height;
enc->fps_num = cfg->fps_num;
enc->fps_den = cfg->fps_den;
enc->input_format = NV_ENC_BUFFER_FORMAT_NV12;
pthread_mutex_init(&enc->registered_mu, NULL);
int rc = 0;
/* 1) Open session с CUDA context. */
NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS sp = { 0 };
sp.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
sp.device = cfg->cuda_ctx;
sp.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
sp.apiVersion = NVENCAPI_VERSION;
NVE_CHECK(g_nvenc_funcs.nvEncOpenEncodeSessionEx(&sp, &enc->session),
fail_alloc);
/* 2) Получить дефолтный config от preset'а. NVENC сам подберёт оптимальные
* параметры для выбранного preset'а + tuning info. */
const GUID *preset_guid = select_preset(cfg->preset);
NV_ENC_PRESET_CONFIG preset_cfg = { 0 };
preset_cfg.version = NV_ENC_PRESET_CONFIG_VER;
preset_cfg.presetCfg.version = NV_ENC_CONFIG_VER;
NVE_CHECK(g_nvenc_funcs.nvEncGetEncodePresetConfigEx(
enc->session, nv_codec_h264, *preset_guid,
NV_ENC_TUNING_INFO_LOW_LATENCY, &preset_cfg),
fail_session);
NV_ENC_CONFIG ec = preset_cfg.presetCfg;
/* 3) Override config: bitrate, GOP, B-frames, repeat SPS/PPS. */
ec.gopLength = cfg->gop_size > 0 ? cfg->gop_size : 25;
ec.frameIntervalP = cfg->num_b_frames + 1; /* 1 = только P, 2 = 1 B, etc */
ec.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR;
ec.rcParams.averageBitRate = cfg->bitrate_kbps * 1000;
ec.rcParams.maxBitRate = cfg->bitrate_kbps * 1000;
ec.rcParams.vbvBufferSize = cfg->bitrate_kbps * 1000;
ec.rcParams.vbvInitialDelay = cfg->bitrate_kbps * 1000;
/* Repeat SPS/PPS перед каждым IDR — важно для RTSP/MP4 где decoder
* может присоединиться к потоку в любой момент. */
ec.encodeCodecConfig.h264Config.repeatSPSPPS = 1;
ec.encodeCodecConfig.h264Config.idrPeriod = ec.gopLength;
/* 4) Init encoder. */
NV_ENC_INITIALIZE_PARAMS ip = { 0 };
ip.version = NV_ENC_INITIALIZE_PARAMS_VER;
ip.encodeGUID = nv_codec_h264;
ip.presetGUID = *preset_guid;
ip.tuningInfo = NV_ENC_TUNING_INFO_LOW_LATENCY;
ip.encodeWidth = cfg->width;
ip.encodeHeight = cfg->height;
ip.darWidth = cfg->width;
ip.darHeight = cfg->height;
ip.frameRateNum = cfg->fps_num;
ip.frameRateDen = cfg->fps_den;
ip.enablePTD = 1; /* Picture type decision — пусть NVENC сам решает */
ip.encodeConfig = &ec;
NVE_CHECK(g_nvenc_funcs.nvEncInitializeEncoder(enc->session, &ip),
fail_session);
/* 5) Создать output bitstream buffer. Phase 1 — один. */
NV_ENC_CREATE_BITSTREAM_BUFFER cb = { 0 };
cb.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
NVE_CHECK(g_nvenc_funcs.nvEncCreateBitstreamBuffer(enc->session, &cb),
fail_session);
enc->output_bitstream = cb.bitstreamBuffer;
*out = enc;
return 0;
fail_session:
g_nvenc_funcs.nvEncDestroyEncoder(enc->session);
fail_alloc:
pthread_mutex_destroy(&enc->registered_mu);
free(enc);
return rc;
}
int cfc_encoder_encode_frame(cfc_encoder_t *enc, CUdeviceptr ptr, int pitch,
int64_t pts_ns,
cfc_encoder_output_cb cb, void *user)
{
if (!enc) return -1;
NV_ENC_REGISTERED_PTR regptr = get_or_register(enc, ptr, pitch);
if (!regptr) return -1;
NV_ENC_MAP_INPUT_RESOURCE mp = { 0 };
mp.version = NV_ENC_MAP_INPUT_RESOURCE_VER;
mp.registeredResource = regptr;
NVENCSTATUS s = g_nvenc_funcs.nvEncMapInputResource(enc->session, &mp);
if (s != NV_ENC_SUCCESS) {
fprintf(stderr, "[cfc/nvenc] nvEncMapInputResource failed: %s\n",
cfc_nvenc_status_str(s));
return -1;
}
NV_ENC_PIC_PARAMS pp = { 0 };
pp.version = NV_ENC_PIC_PARAMS_VER;
pp.inputBuffer = mp.mappedResource;
pp.outputBitstream = enc->output_bitstream;
pp.bufferFmt = mp.mappedBufferFmt;
pp.inputWidth = enc->width;
pp.inputHeight = enc->height;
pp.inputPitch = pitch;
pp.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
pp.inputTimeStamp = (uint64_t)pts_ns;
s = g_nvenc_funcs.nvEncEncodePicture(enc->session, &pp);
int rc = 0;
if (s == NV_ENC_ERR_NEED_MORE_INPUT) {
/* OK, NVENC буферизует — выходного кадра пока нет. */
} else if (s == NV_ENC_SUCCESS) {
/* Есть готовый закодированный кадр — вытащить через callback. */
if (drain_output(enc, cb, user) < 0) rc = -1;
} else {
fprintf(stderr, "[cfc/nvenc] nvEncEncodePicture failed: %s\n",
cfc_nvenc_status_str(s));
rc = -1;
}
NVENCSTATUS us = g_nvenc_funcs.nvEncUnmapInputResource(enc->session,
mp.mappedResource);
if (us != NV_ENC_SUCCESS) {
fprintf(stderr, "[cfc/nvenc] nvEncUnmapInputResource failed: %s\n",
cfc_nvenc_status_str(us));
rc = -1;
}
return rc;
}
int cfc_encoder_flush(cfc_encoder_t *enc, cfc_encoder_output_cb cb, void *user)
{
if (!enc) return -1;
/* Send EOS signal. */
NV_ENC_PIC_PARAMS pp = { 0 };
pp.version = NV_ENC_PIC_PARAMS_VER;
pp.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
NVENCSTATUS s = g_nvenc_funcs.nvEncEncodePicture(enc->session, &pp);
if (s != NV_ENC_SUCCESS && s != NV_ENC_ERR_NEED_MORE_INPUT) {
fprintf(stderr, "[cfc/nvenc] flush: nvEncEncodePicture(EOS) failed: %s\n",
cfc_nvenc_status_str(s));
return -1;
}
/* Вытащить все буферизованные кадры. */
int got;
do {
got = drain_output(enc, cb, user);
} while (got > 0);
return got < 0 ? -1 : 0;
}
int cfc_encoder_get_sequence_params(cfc_encoder_t *enc,
uint8_t *out, size_t *inout_size)
{
if (!enc || !out || !inout_size) return -1;
NV_ENC_SEQUENCE_PARAM_PAYLOAD spp = { 0 };
spp.version = NV_ENC_SEQUENCE_PARAM_PAYLOAD_VER;
spp.inBufferSize = (uint32_t)*inout_size;
spp.spsppsBuffer = out;
uint32_t written = 0;
spp.outSPSPPSPayloadSize = &written;
NVENCSTATUS s = g_nvenc_funcs.nvEncGetSequenceParams(enc->session, &spp);
if (s != NV_ENC_SUCCESS) {
fprintf(stderr, "[cfc/nvenc] nvEncGetSequenceParams failed: %s\n",
cfc_nvenc_status_str(s));
return -1;
}
*inout_size = written;
return 0;
}
int cfc_encoder_destroy(cfc_encoder_t *enc)
{
if (!enc) return 0;
/* Unregister всех зарегистрированных входных буферов. */
pthread_mutex_lock(&enc->registered_mu);
for (int i = 0; i < enc->registered_count; i++) {
g_nvenc_funcs.nvEncUnregisterResource(enc->session,
enc->registered[i].regptr);
}
enc->registered_count = 0;
pthread_mutex_unlock(&enc->registered_mu);
pthread_mutex_destroy(&enc->registered_mu);
if (enc->output_bitstream) {
g_nvenc_funcs.nvEncDestroyBitstreamBuffer(enc->session,
enc->output_bitstream);
}
if (enc->session) {
g_nvenc_funcs.nvEncDestroyEncoder(enc->session);
}
free(enc);
return 0;
}
+109
View File
@@ -0,0 +1,109 @@
/* Динамическая загрузка libnvidia-encode.so и инициализация таблицы
* NVENC API. См. nvenc_loader.h для контракта.
*
* Lib path: libnvidia-encode.so.1 (стандартное имя в Linux x86_64,
* поставляется NVIDIA-драйвером в /usr/lib/x86_64-linux-gnu/).
*
* API version: используем макрос NVENCAPI_VERSION из nv-codec-headers.
* SDK сам проверит совместимость драйвера (вернёт NV_ENC_ERR_INVALID_VERSION
* если драйвер старше чем заголовки требуют).
*/
#include "nvenc_loader.h"
#include <dlfcn.h>
#include <pthread.h>
#include <stdio.h>
#include <string.h>
NV_ENCODE_API_FUNCTION_LIST g_nvenc_funcs;
static pthread_once_t g_init_once = PTHREAD_ONCE_INIT;
static int g_init_result = -1;
/* Сигнатура единственного символа который мы dlsym'им из libnvidia-encode.so.
* Все остальные функции SDK достаются через NvEncodeAPICreateInstance, который
* заполняет таблицу указателями. */
typedef NVENCSTATUS (NVENCAPI *NvEncodeAPICreateInstance_fn)(NV_ENCODE_API_FUNCTION_LIST *);
static void load_once(void)
{
/* RTLD_LAZY — функции резолвятся по мере вызова, что нормально для
* подсистемы которую мы используем целиком. RTLD_NODELETE — не выгружать
* lib при dlclose (мы и так держим до выхода процесса). */
void *handle = dlopen("libnvidia-encode.so.1", RTLD_LAZY | RTLD_NODELETE);
if (!handle) {
/* Fallback на имя без версии (некоторые системы держат symlink). */
handle = dlopen("libnvidia-encode.so", RTLD_LAZY | RTLD_NODELETE);
}
if (!handle) {
fprintf(stderr,
"[cfc/nvenc] dlopen libnvidia-encode.so.1 failed: %s\n"
" → NVIDIA-драйвер не установлен либо libnvidia-encode не в LD_LIBRARY_PATH\n",
dlerror());
return;
}
NvEncodeAPICreateInstance_fn create_instance =
(NvEncodeAPICreateInstance_fn)dlsym(handle, "NvEncodeAPICreateInstance");
if (!create_instance) {
fprintf(stderr,
"[cfc/nvenc] dlsym NvEncodeAPICreateInstance failed: %s\n"
" → libnvidia-encode.so есть, но без ожидаемого символа\n",
dlerror());
return;
}
memset(&g_nvenc_funcs, 0, sizeof(g_nvenc_funcs));
g_nvenc_funcs.version = NV_ENCODE_API_FUNCTION_LIST_VER;
NVENCSTATUS s = create_instance(&g_nvenc_funcs);
if (s != NV_ENC_SUCCESS) {
fprintf(stderr,
"[cfc/nvenc] NvEncodeAPICreateInstance failed: %s (status %d)\n"
" → версия драйвера несовместима с NVENC API %u.%u\n",
cfc_nvenc_status_str(s), s,
NVENCAPI_MAJOR_VERSION, NVENCAPI_MINOR_VERSION);
return;
}
g_init_result = 0;
}
int cfc_nvenc_loader_init(void)
{
pthread_once(&g_init_once, load_once);
return g_init_result;
}
const char *cfc_nvenc_status_str(NVENCSTATUS s)
{
switch (s) {
case NV_ENC_SUCCESS: return "NV_ENC_SUCCESS";
case NV_ENC_ERR_NO_ENCODE_DEVICE: return "NV_ENC_ERR_NO_ENCODE_DEVICE";
case NV_ENC_ERR_UNSUPPORTED_DEVICE: return "NV_ENC_ERR_UNSUPPORTED_DEVICE";
case NV_ENC_ERR_INVALID_ENCODERDEVICE: return "NV_ENC_ERR_INVALID_ENCODERDEVICE";
case NV_ENC_ERR_INVALID_DEVICE: return "NV_ENC_ERR_INVALID_DEVICE";
case NV_ENC_ERR_DEVICE_NOT_EXIST: return "NV_ENC_ERR_DEVICE_NOT_EXIST";
case NV_ENC_ERR_INVALID_PTR: return "NV_ENC_ERR_INVALID_PTR";
case NV_ENC_ERR_INVALID_EVENT: return "NV_ENC_ERR_INVALID_EVENT";
case NV_ENC_ERR_INVALID_PARAM: return "NV_ENC_ERR_INVALID_PARAM";
case NV_ENC_ERR_INVALID_CALL: return "NV_ENC_ERR_INVALID_CALL";
case NV_ENC_ERR_OUT_OF_MEMORY: return "NV_ENC_ERR_OUT_OF_MEMORY";
case NV_ENC_ERR_ENCODER_NOT_INITIALIZED: return "NV_ENC_ERR_ENCODER_NOT_INITIALIZED";
case NV_ENC_ERR_UNSUPPORTED_PARAM: return "NV_ENC_ERR_UNSUPPORTED_PARAM";
case NV_ENC_ERR_LOCK_BUSY: return "NV_ENC_ERR_LOCK_BUSY";
case NV_ENC_ERR_NOT_ENOUGH_BUFFER: return "NV_ENC_ERR_NOT_ENOUGH_BUFFER";
case NV_ENC_ERR_INVALID_VERSION: return "NV_ENC_ERR_INVALID_VERSION";
case NV_ENC_ERR_MAP_FAILED: return "NV_ENC_ERR_MAP_FAILED";
case NV_ENC_ERR_NEED_MORE_INPUT: return "NV_ENC_ERR_NEED_MORE_INPUT";
case NV_ENC_ERR_ENCODER_BUSY: return "NV_ENC_ERR_ENCODER_BUSY";
case NV_ENC_ERR_EVENT_NOT_REGISTERD: return "NV_ENC_ERR_EVENT_NOT_REGISTERD";
case NV_ENC_ERR_GENERIC: return "NV_ENC_ERR_GENERIC";
case NV_ENC_ERR_INCOMPATIBLE_CLIENT_KEY: return "NV_ENC_ERR_INCOMPATIBLE_CLIENT_KEY";
case NV_ENC_ERR_UNIMPLEMENTED: return "NV_ENC_ERR_UNIMPLEMENTED";
case NV_ENC_ERR_RESOURCE_REGISTER_FAILED: return "NV_ENC_ERR_RESOURCE_REGISTER_FAILED";
case NV_ENC_ERR_RESOURCE_NOT_REGISTERED: return "NV_ENC_ERR_RESOURCE_NOT_REGISTERED";
case NV_ENC_ERR_RESOURCE_NOT_MAPPED: return "NV_ENC_ERR_RESOURCE_NOT_MAPPED";
default: return "NV_ENC_ERR_UNKNOWN";
}
}
+29
View File
@@ -0,0 +1,29 @@
/* Внутренний заголовок — динамическая загрузка libnvidia-encode.so
* через dlopen, чтобы не линковаться статически с проприетарным SDK.
*
* Загружает символ `NvEncodeAPICreateInstance`, вызывает его для
* наполнения NV_ENCODE_API_FUNCTION_LIST — таблицы указателей на все
* функции NVENC. Дальше весь код в src/nvenc.c вызывает функции через
* эту таблицу.
*/
#ifndef CFC_NVENC_LOADER_H
#define CFC_NVENC_LOADER_H
#include <ffnvcodec/nvEncodeAPI.h>
/* Глобально кешированная таблица функций NVENC. После первого успешного
* cfc_nvenc_loader_init() остаётся валидной до выхода процесса. */
extern NV_ENCODE_API_FUNCTION_LIST g_nvenc_funcs;
/* Загружает libnvidia-encode.so и наполняет g_nvenc_funcs.
* Возвращает 0 при успехе, -1 если SDK не найден или версия несовместима.
* Идемпотентно — повторные вызовы дешёвые (возвращают 0 если уже загружено).
*
* Thread-safe для первого вызова через pthread_once. */
int cfc_nvenc_loader_init(void);
/* Текстовое описание ошибки NVENC SDK по коду NVENCSTATUS. */
const char *cfc_nvenc_status_str(NVENCSTATUS s);
#endif /* CFC_NVENC_LOADER_H */
+294
View File
@@ -0,0 +1,294 @@
/* Обвязка над cuframes_subscriber. Реализация публичного API
* cuframes_composer/source.h.
*
* Архитектура:
* Отдельный поток на источник. cuframes_subscriber_next блокирующий, поток
* циклически зовёт его с таймаутом. На каждый успешный кадр — обновляет
* snapshot (под mutex'ом). На таймаут — переходит к проверке состояния.
*
* Snapshot pattern:
* Главный поток (тот что зовёт get_latest) не блокируется на subscribe.
* Он читает under mutex последний сохранённый CUdeviceptr + meta.
* Mutex короткий (просто копия указателя и нескольких int'ов).
*
* State machine:
* DISCONNECTED → создаём подписку → CONNECTING
* CONNECTING → подписка успешна, первый кадр получен → ACTIVE
* → подписка fail → DEAD (ждём reconnect_backoff)
* ACTIVE → последний кадр < stale_threshold_ms → остаёмся ACTIVE
* → > stale_threshold_ms → STALE
* STALE → новый кадр → ACTIVE
* → нет > dead_threshold_ms → DEAD (destroy subscriber)
* DEAD → ждём backoff (exp от reconnect_min до reconnect_max) → CONNECTING
*
* Phase 1 особенности:
* - Один источник на cfc_source_t (нет multi-source агрегации, это compose).
* - cuframes_subscriber_release вызывается при следующем get_latest либо
* при выходе. Это означает что caller не может «удерживать» snapshot
* дольше чем до следующего get_latest — должен прочитать сразу.
* В Phase 2 это уточнится при переходе на double buffering.
*/
#include "../include/cuframes_composer/source.h"
#include <cuframes/cuframes.h>
#include <errno.h>
#include <pthread.h>
#include <stdatomic.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
/* Внутренний таймаут блокирующего cuframes_subscriber_next. Короткий
* чтобы поток мог периодически проверять stop_flag и состояние. */
#define CFC_SOURCE_NEXT_TIMEOUT_MS 200
struct cfc_source {
cfc_source_config_t cfg;
char key_copy[64]; /* персистентная копия cfg.key */
char name_copy[32]; /* персистентная копия cfg.consumer_name */
pthread_t thread;
int thread_started;
_Atomic int stop_flag;
pthread_mutex_t state_mu;
cfc_source_state_t state;
/* Snapshot — обновляется потоком, читается через get_latest. */
cuframes_subscriber_t *sub; /* nullable — есть только в ACTIVE/STALE */
cuframes_frame_t *current_frame; /* удерживаемый frame; release при следующем next() */
cfc_source_snapshot_t snapshot;
int64_t last_frame_us; /* CLOCK_MONOTONIC момент последнего успешного кадра */
};
static int64_t now_us(void)
{
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return (int64_t)ts.tv_sec * 1000000 + ts.tv_nsec / 1000;
}
static void set_state(cfc_source_t *src, cfc_source_state_t s)
{
pthread_mutex_lock(&src->state_mu);
if (src->state != s) {
src->state = s;
src->snapshot.state = s;
}
pthread_mutex_unlock(&src->state_mu);
}
static void release_current_frame(cfc_source_t *src)
{
if (src->current_frame && src->sub) {
cuframes_subscriber_release(src->sub, src->current_frame);
src->current_frame = NULL;
}
}
static int try_subscribe(cfc_source_t *src)
{
cuframes_subscriber_config_t scfg = { 0 };
scfg.key = src->key_copy;
scfg.consumer_name = src->name_copy;
scfg.mode = CUFRAMES_MODE_NEWEST_ONLY;
scfg.cuda_device = src->cfg.cuda_device;
scfg.connect_timeout_ms = 2000;
int r = cuframes_subscriber_create(&scfg, &src->sub);
if (r != CUFRAMES_OK) {
fprintf(stderr,
"[cfc/source:%s] subscriber_create failed: %s\n",
src->name_copy, cuframes_strerror(r));
return -1;
}
return 0;
}
static void destroy_subscriber(cfc_source_t *src)
{
release_current_frame(src);
if (src->sub) {
cuframes_subscriber_destroy(src->sub);
src->sub = NULL;
}
}
/* Основной поток. Цикл: subscribe → next → update snapshot → проверка
* stale/dead → reconnect при необходимости. */
static void *source_thread(void *arg)
{
cfc_source_t *src = (cfc_source_t *)arg;
int64_t reconnect_backoff_ms = src->cfg.reconnect_min_ms;
while (!atomic_load(&src->stop_flag)) {
cfc_source_state_t cur;
pthread_mutex_lock(&src->state_mu);
cur = src->state;
pthread_mutex_unlock(&src->state_mu);
switch (cur) {
case CFC_SOURCE_DISCONNECTED:
case CFC_SOURCE_DEAD: {
/* Ждём backoff либо stop. */
int64_t wait_ms = reconnect_backoff_ms;
while (wait_ms > 0 && !atomic_load(&src->stop_flag)) {
int chunk = wait_ms > 100 ? 100 : (int)wait_ms;
struct timespec ts = {.tv_sec = chunk / 1000,
.tv_nsec = (long)(chunk % 1000) * 1000000L};
nanosleep(&ts, NULL);
wait_ms -= chunk;
}
if (atomic_load(&src->stop_flag)) break;
set_state(src, CFC_SOURCE_CONNECTING);
break;
}
case CFC_SOURCE_CONNECTING: {
if (try_subscribe(src) == 0) {
set_state(src, CFC_SOURCE_ACTIVE);
reconnect_backoff_ms = src->cfg.reconnect_min_ms; /* сброс backoff */
} else {
set_state(src, CFC_SOURCE_DEAD);
/* Удвоить backoff до max. */
reconnect_backoff_ms *= 2;
if (reconnect_backoff_ms > src->cfg.reconnect_max_ms) {
reconnect_backoff_ms = src->cfg.reconnect_max_ms;
}
}
break;
}
case CFC_SOURCE_ACTIVE:
case CFC_SOURCE_STALE: {
cuframes_frame_t *frame = NULL;
int r = cuframes_subscriber_next(src->sub, NULL, &frame,
CFC_SOURCE_NEXT_TIMEOUT_MS);
if (r == CUFRAMES_OK) {
/* Освобождаем предыдущий кадр (если был удержан caller'ом
* во время предыдущего snapshot — поздно, но это Phase 1
* упрощение, см. описание сверху). */
release_current_frame(src);
src->current_frame = frame;
/* Обновляем snapshot под mutex'ом. */
int32_t w = 0, h = 0;
cuframes_frame_size(frame, &w, &h);
pthread_mutex_lock(&src->state_mu);
src->snapshot.ptr = (CUdeviceptr)cuframes_frame_cuda_ptr(frame);
src->snapshot.width = w;
src->snapshot.height = h;
src->snapshot.pitch_y = cuframes_frame_pitch_y(frame);
src->snapshot.pitch_uv = cuframes_frame_pitch_uv(frame);
src->snapshot.pts_ns = cuframes_frame_pts_ns(frame);
src->snapshot.seq = cuframes_frame_seq(frame);
src->last_frame_us = now_us();
src->snapshot.last_frame_age_us = 0;
if (src->state == CFC_SOURCE_STALE) {
src->state = CFC_SOURCE_ACTIVE;
src->snapshot.state = CFC_SOURCE_ACTIVE;
}
pthread_mutex_unlock(&src->state_mu);
} else if (r == CUFRAMES_ERR_TIMEOUT || r == CUFRAMES_ERR_WOULD_BLOCK) {
/* Нет нового кадра — проверим age, может быть STALE/DEAD. */
int64_t age_ms = (now_us() - src->last_frame_us) / 1000;
if (age_ms > src->cfg.dead_threshold_ms) {
fprintf(stderr,
"[cfc/source:%s] no frame for %lldms → DEAD\n",
src->name_copy, (long long)age_ms);
destroy_subscriber(src);
set_state(src, CFC_SOURCE_DEAD);
} else if (age_ms > src->cfg.stale_threshold_ms) {
set_state(src, CFC_SOURCE_STALE);
}
} else if (r == CUFRAMES_ERR_DISCONNECTED) {
fprintf(stderr,
"[cfc/source:%s] DISCONNECTED from publisher\n",
src->name_copy);
destroy_subscriber(src);
set_state(src, CFC_SOURCE_DEAD);
} else {
fprintf(stderr,
"[cfc/source:%s] cuframes_subscriber_next failed: %s\n",
src->name_copy, cuframes_strerror(r));
destroy_subscriber(src);
set_state(src, CFC_SOURCE_DEAD);
}
break;
}
}
}
destroy_subscriber(src);
return NULL;
}
/* ── Public API ───────────────────────────────────────────────────────── */
int cfc_source_create(const cfc_source_config_t *cfg, cfc_source_t **out)
{
if (!cfg || !cfg->key || !cfg->consumer_name || !out) return -1;
cfc_source_t *src = calloc(1, sizeof(*src));
if (!src) return -1;
src->cfg = *cfg;
strncpy(src->key_copy, cfg->key, sizeof(src->key_copy) - 1);
strncpy(src->name_copy, cfg->consumer_name, sizeof(src->name_copy) - 1);
src->cfg.key = src->key_copy;
src->cfg.consumer_name = src->name_copy;
/* Дефолты */
if (src->cfg.reconnect_min_ms <= 0) src->cfg.reconnect_min_ms = 1000;
if (src->cfg.reconnect_max_ms <= 0) src->cfg.reconnect_max_ms = 30000;
if (src->cfg.stale_threshold_ms <= 0) src->cfg.stale_threshold_ms = 500;
if (src->cfg.dead_threshold_ms <= 0) src->cfg.dead_threshold_ms = 5000;
pthread_mutex_init(&src->state_mu, NULL);
src->state = CFC_SOURCE_DISCONNECTED;
src->snapshot.state = CFC_SOURCE_DISCONNECTED;
src->snapshot.last_frame_age_us = -1;
atomic_init(&src->stop_flag, 0);
if (pthread_create(&src->thread, NULL, source_thread, src) != 0) {
pthread_mutex_destroy(&src->state_mu);
free(src);
return -1;
}
src->thread_started = 1;
*out = src;
return 0;
}
int cfc_source_get_latest(cfc_source_t *src, cfc_source_snapshot_t *out)
{
if (!src || !out) return -1;
pthread_mutex_lock(&src->state_mu);
*out = src->snapshot;
if (src->last_frame_us > 0) {
out->last_frame_age_us = now_us() - src->last_frame_us;
} else {
out->last_frame_age_us = -1;
}
pthread_mutex_unlock(&src->state_mu);
return 0;
}
int cfc_source_destroy(cfc_source_t *src)
{
if (!src) return 0;
atomic_store(&src->stop_flag, 1);
if (src->thread_started) {
pthread_join(src->thread, NULL);
}
pthread_mutex_destroy(&src->state_mu);
free(src);
return 0;
}
Vendored Submodule
+1
Submodule third_party/cuframes added at 655649f4d8