Releases: nvidia-holoscan/holoscan-sdk
v2.4.0
Release Artifacts
- 🐋 Docker container: tag
v2.4.0-dgpu
andv2.4.0-igpu
- 🐍 Python wheel:
pip install holoscan==2.4.0
- 📦️ Debian packages:
2.4.0.1-1
- 📕 Documentation
See supported platforms for compatibility.
Release Notes
New Features and Improvements
-
The Holoscan CLI packager has been updated to create application containers that are up to 78% smaller than their previous standard container based size. The new
--includes
option allows the packager to include only runtime dependencies relevant to the application. Refer to the documentation for more information. -
The Holoscan pipeline metadata feature introduced for the C++ API in release v2.3 is now also available from the Python API. Interaction with metadata can be done with an API very similar to Python's built-in dictionaries. Please see the dynamic application metadata section of the user guide for details.
-
The V4L example now supports a YUV configuration to display YUV input directly without RGB conversion.
-
Added support to build Holoscan SDK without docker cache in
run
script e.g.,./run build --no-cache
Core
-
It is now possible to set a
HOLOSCAN_QUEUE_POLICY
environment variable to override the default queue policy that is used by the input and output ports of the SDK. Valid options (case insensitive) are:- "pop": a new item that arrives when the queue is full replaces the oldest item
- "reject": a new item that arrives when the queue is discarded
- "fail": terminate the application if a new item arrives when the queue is full
The default behavior remains "fail" if the environment variable is not specified. If an operator's
setup
method explicitly sets a receiver or transmitter via theIOSpec::connector
method, this default value does not override the policy of that connector. -
Improve the handling of extra arguments passed to the application via CLI.
- Now calls the
allow_extras()
method to permit extra arguments instead of ignoring them on anExtrasError
in the CLI11 parser.
- Now calls the
-
The SDK no longer spawns a new process to check for unused network ports for UCX communication when running a distributed application. Previously, this process caused issues such as redundant system resource consumption when the Holoscan distributed application was run as part of a larger application (e.g., importing Holoscan as a Python module after importing other modules) because a new process was created solely for checking unused network ports. This issue has been resolved by performing the network port check in in-process mode.
Operators/Resources
- Two new operators useful for examples and testing were added.
PingTensorTxOp
will emit a TensorMap containing a single tensor with user-specified name, shape, data type and storage type (e.g. host vs. device).PingTensorRxOp
will receive a message containing a TensorMap and print some attributes of any tensors contained within it. Versions of these previously existed in examples and test code, but have now been moved to a common public location (holoscan::ops
namespace for C++ and underholoscan.operators
for Python). - The
V4L2VideoCaptureOp
now supports passing the input buffer unmodified to the output. This can be enabled by using the parameterpass_through
, by default this is disabled. - Handle/enhance various cases of multi-receiver input ports (
holoscan::InputContext::receive<std::vector<T>>()
)- Support receiving an array of TensorMap items from the input port.
- Improve handling of cases where no data or null pointers are received from the input port.
- Throw an invalid argument exception if the operator attempts to receive non-vector data (
op_input.receive<T>()
) from an input port with a queue size ofIOSpec::kAnySize
. - Avoid using
nvidia::TypenameAsString
for the type name in error messages, as it may include characters that are not permitted in the message (e.g.,{anonymous}
), which could be interpreted as a format specifier. This can result in an exception being thrown during message formatting.
- The
HolovizOp
now supports YUV (aka YCbCr) images as input. Various420
and422
formats are supported.- New image formats:
y8u8y8v8_422_unorm
u8y8v8y8_422_unorm
y8_u8v8_2plane_420_unorm
y8_u8v8_2plane_422_unorm
y8_u8_v8_3plane_420_unorm
y8_u8_v8_3plane_422_unorm
y16_u16v16_2plane_420_unorm
y16_u16v16_2plane_422_unorm
y16_u16_v16_3plane_420_unorm
y16_u16_v16_3plane_422_unorm
- YUV color model conversions:
yuv_601
yuv_709
yuv_2020
- YUV ranges:
itu_full
itu_narrow
- Chroma locations in x and y:
cosited_even
midpoint
- New image formats:
- The
AJASourceOp
now supports the following video formats:- 720p @ 50, 59.94, 60Hz
- 1080i @ 50, 59.94, 60Hz
- 1080p @ 23.98, 24, 25, 29.97, 30, 50, 59.94, 60Hz
- 3840x2160 (UHD) @ 23.98, 24, 25, 29.97, 30, 50, 59.94, 60Hz
- 4096x2160 (4K) @ 23.98, 24, 25, 29.97, 30, 50, 59.94, 60Hz
Holoviz module
- Now supports YUV (aka YCbCr) images and YUV conversion parameters. The functions to specify image layer data have been extended to support planar formats.
- New entry point
ImageYuvModelConversion()
to specify the YUV model conversion (BT.601, BT.709, BT.2020) - New entry point
ImageYuvRange()
to specify the YUV range (ITU full and ITU narrow) - New entry point
ImageChromaLocation()
to specify the chroma location (cosited even and midpoint)
- New entry point
Utils
- An
aja_build.sh
script was added to automate the download, build, and loading of the AJA NTV2 drivers and SDK.
HoloHub
Documentation
Breaking Changes
Bug fixes
Issue | Description |
---|---|
- | Holoviz operator fails with Surface format '29, 0' not supported when enabling sRGB framebuffer in headless mode. |
- | Fixed a bug where the run vscode --parallel <num_workers> command was not working as expected, displaying the message arg: unbound variable . |
4791938 | v4l_camera doesn't work with the USB camera when 800x600 is set, and there are multiple sizes available for width 800. |
4792457 | Heap memory error was found in GXFParameterAdaptor with AddressSanitizer (ASAN) during dynamic analysis. |
4752615 | In Python, the operator's parameter values are not available in the initialize() method. This bug was introduced in version 2.1.0. |
4510522 | V4L2VideoCaptureOp does not work with RGB. |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
4062979 | When Operators connected in a Directed Acyclic Graph (DAG) are executed in a multithreaded scheduler, it is not ensured that their execution order in the graph is adhered. |
4267272 | AJA drivers cannot be built with RDMA on IGX SW 1.0 DP iGPU due to missing nv-p2p.h . Expected to be addressed in IGX SW 1.0 GA. |
4384768 | No RDMA support on JetPack 6.0 DP and IGX SW 1.0 DP iGPU due to missing nv-p2p kernel module. Expected to be addressed in JP 6.0 GA and IGX SW 1.0 GA respectively. |
4190019 | Holoviz segfaults on multi-gpu setup when specifying device using the --gpus flag with docker run . Current workaround is to use CUDA_VISIBLE_DEVICES in the container instead. |
4210082 | v4l_camera example seg faults at exit. |
4339399 | High CPU usage observed with video_replayer_distributed application. While the high CPU usage associated with the GXF UCX extension has been fixed since v1.0, distributed applications using the MultiThreadScheduler (with the check_recession_period_ms parameter set to 0 by default) may still experience high CPU usage. Setting the HOLOSCAN_CHECK_RECESSION_PERIOD_MS environment variable to a value greater than 0 (e.g. 1.5 ) can help reduce CPU usage. However, this may result in increased latency for the application until the MultiThreadScheduler switches to an event-based multithreaded scheduler. |
4318442 | UCX cuda_ipc protocol doesn't work in Docker containers on x86_64. As a workaround, we are currently disabling the UCX cuda_ipc protocol on all platforms via the UCX_TLS environment variable. |
4325468 | The V4L2VideoCapture operator only supports YUYV and AB24 source pixel formats, and only outputs the RGBA GXF video format. Other source pixel formats compatible with V4L2 can be manually defined by the user, but they're assumed to be equivalent to RGBA8888. |
4325585 | Applications using MultiThreadScheduler may exit early due to timeouts. This occurs when the stop_on_deadlock_timeout parameter is improperly set to a value equal to or less than check_recession_period_ms , particularly if check_recession_period_ms is greater than zero. |
4301203 | HDMI IN fails in v4l2_camera on IGX Orin Devkit for some resolution or formats. Try the latest firmware as a partial fix. Driver-level fixes expected in IGX SW 1.0 GA. |
4384348 | UCX termination (either ctrl+c , press 'Esc' or clicking close button) is not smooth and can show multiple error messages. |
4481171 | Running the driver for a distributed applications on IGX Orin devkits fails when connected to other systems through eth1. A workar... |
v2.3.0
Release Artifacts
- 🐋 Docker container: tag
v2.3.0-dgpu
andv2.3.0-igpu
- 🐍 Python wheel:
pip install holoscan==2.3.0
- 📦️ Debian packages:
2.3.0.1-1
- 📕 Documentation
See supported platforms for compatibility.
Release Notes
New Features and Improvements
Core
- Explicitly delete the copy constructor and assignment operator of Config/Executor/Graph classes private to prevent copying of these objects. This change ensures that the Config/Executor/Graph classes are not copied, as they are not intended to be copied, which is inefficient. This change is backward compatible, as the classes are still movable.
- [Internal]
Application::fragment_graph_
is now astd::shared_ptr
toFragmentGraph
to prevent copying ofFragmentGraph
objects in Python bindings. - [Internal]
Fragment::graph_
is now astd::shared_ptr
toOperatorGraph
to prevent copying ofOperatorGraph
objects in Python bindings. - [Internal] Added
Fragment::config_shared()
/Fragment::executor_shared()
/Fragment::graph_shared()
to return a shared pointer to theConfig
/Executor
/OperatorGraph
objects respectively.
- [Internal]
Operators/Resources
- A new decorator,
holoscan.decorator.create_op
is provided that can wrap an existing function or generator as a native Python Operator. This new API is still considered experimental and may be updated in a subsequent release based on initial feedback. An example of using this decorator is provided underexamples/python_decorator/video_replayer.py
as well as in the test applications withinpython/tests/system/test_decorator_apps.py
.
Utils
HoloHub
Documentation
Breaking Changes
Bug fixes
Issue | Description |
---|---|
4687735 | Fixed a bug where the CountCondition in Python doesn't accept a negative value for the count parameter, even though it is allowed in C++. Note that using a negative value for count is not recommended as it would lead to an infinite loop until it reaches the minimum value of the data type (int64_t) and then starts counting down from the maximum value up to zero. |
4689604 | Fixed a bug where GXF extensions listed in the config file (YAML) are not loaded when GXFCodeletOp or GXFComponentResource is used. Test cases have been added to verify the fix, and the documentation has been updated to reflect the changes. The HOLOSCAN_WRAP_GXF_COMPONENT_AS_RESOURCE macro has also been updated to support the constructor with no arguments. |
4706559 | Fixed a bug where nullptr was returned even when no message was available in calls to holoscan::IOContext::receive<T*>() or holoscan::IOContext::receive<std::shared_ptr<T>> in the C++ API. Now, the receive method correctly returns a holoscan::unexpected<holoscan::RuntimeError> value when no message is available. |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
4062979 | When Operators connected in a Directed Acyclic Graph (DAG) are executed in a multithreaded scheduler, it is not ensured that their execution order in the graph is adhered. |
4267272 | AJA drivers cannot be built with RDMA on IGX SW 1.0 DP iGPU due to missing nv-p2p.h . Expected to be addressed in IGX SW 1.0 GA. |
4384768 | No RDMA support on JetPack 6.0 DP and IGX SW 1.0 DP iGPU due to missing nv-p2p kernel module. Expected to be addressed in JP 6.0 GA and IGX SW 1.0 GA respectively. |
4190019 | Holoviz segfaults on multi-gpu setup when specifying device using the --gpus flag with docker run . Current workaround is to use CUDA_VISIBLE_DEVICES in the container instead. |
4210082 | v4l_camera example seg faults at exit. |
4339399 | High CPU usage observed with video_replayer_distributed application. While the high CPU usage associated with the GXF UCX extension has been fixed since v1.0, distributed applications using the MultiThreadScheduler (with the check_recession_period_ms parameter set to 0 by default) may still experience high CPU usage. Setting the HOLOSCAN_CHECK_RECESSION_PERIOD_MS environment variable to a value greater than 0 (e.g. 1.5 ) can help reduce CPU usage. However, this may result in increased latency for the application until the MultiThreadScheduler switches to an event-based multithreaded scheduler. |
4318442 | UCX cuda_ipc protocol doesn't work in Docker containers on x86_64. As a workaround, we are currently disabling the UCX cuda_ipc protocol on all platforms via the UCX_TLS environment variable. |
4325468 | The V4L2VideoCapture operator only supports YUYV and AB24 source pixel formats, and only outputs the RGBA GXF video format. Other source pixel formats compatible with V4L2 can be manually defined by the user, but they're assumed to be equivalent to RGBA8888. |
4325585 | Applications using MultiThreadScheduler may exit early due to timeouts. This occurs when the stop_on_deadlock_timeout parameter is improperly set to a value equal to or less than check_recession_period_ms , particularly if check_recession_period_ms is greater than zero. |
4301203 | HDMI IN fails in v4l2_camera on IGX Orin Devkit for some resolution or formats. Try the latest firmware as a partial fix. Driver-level fixes expected in IGX SW 1.0 GA. |
4384348 | UCX termination (either ctrl+c , press 'Esc' or clicking close button) is not smooth and can show multiple error messages. |
4481171 | Running the driver for a distributed applications on IGX Orin devkits fails when connected to other systems through eth1. A workaround is to use eth0 port to connect to other systems for distributed workloads. |
4458192 | In scenarios where distributed applications have both the driver and workers running on the same host, either within a Docker container or directly on the host, there's a possibility of encountering "Address already in use" errors. A potential solution is to assign a different port number to the HOLOSCAN_HEALTH_CHECK_PORT environment variable (default: 8777 ), for example, by using export HOLOSCAN_HEALTH_CHECK_PORT=8780 . |
Wayland: holoscan::viz::Init() with existing GLFW window fails. |
Holoscan SDK v2.2.0 Release
Release Artifacts
- 🐋 Docker container: tag
v2.2.0-dgpu
andv2.2.0-igpu
- 🐍 Python wheel:
pip install holoscan==2.2.0
- 📦️ Debian packages:
2.2.0.0-1
- 📕 Documentation
See supported platforms for compatibility.
Release Notes
New Features and Improvements
Core
- Explicitly delete the copy constructor and assignment operator of Config/Executor/Graph classes private to prevent copying of these objects. This change ensures that the Config/Executor/Graph classes are not copied, as they are not intended to be copied, which is inefficient. This change is backward compatible, as the classes are still movable.
- [Internal]
Application::fragment_graph_
is now astd::shared_ptr
toFragmentGraph
to prevent copying ofFragmentGraph
objects in Python bindings. - [Internal]
Fragment::graph_
is now astd::shared_ptr
toOperatorGraph
to prevent copying ofOperatorGraph
objects in Python bindings. - [Internal] Added
Fragment::config_shared()
/Fragment::executor_shared()
/Fragment::graph_shared()
to return a shared pointer to theConfig
/Executor
/OperatorGraph
objects respectively.
- [Internal]
Operators/Resources
- A new decorator,
holoscan.decorator.create_op
is provided that can wrap an existing function or generator as a native Python Operator. This new API is still considered experimental and may be updated in a subsequent release based on initial feedback. An example of using this decorator is provided underexamples/python_decorator/video_replayer.py
as well as in the test applications withinpython/tests/system/test_decorator_apps.py
.
Utils
HoloHub
Documentation
Breaking Changes
Bug fixes
Issue | Description |
---|---|
4687735 | Fixed a bug where the CountCondition in Python doesn't accept a negative value for the count parameter, even though it is allowed in C++. Note that using a negative value for count is not recommended as it would lead to an infinite loop until it reaches the minimum value of the data type (int64_t) and then starts counting down from the maximum value up to zero. |
4689604 | Fixed a bug where GXF extensions listed in the config file (YAML) are not loaded when GXFCodeletOp or GXFComponentResource is used. Test cases have been added to verify the fix, and the documentation has been updated to reflect the changes. The HOLOSCAN_WRAP_GXF_COMPONENT_AS_RESOURCE macro has also been updated to support the constructor with no arguments. |
4706559 | Fixed a bug where nullptr was returned even when no message was available in calls to holoscan::IOContext::receive<T*>() or holoscan::IOContext::receive<std::shared_ptr<T>> in the C++ API. Now, the receive method correctly returns a holoscan::unexpected<holoscan::RuntimeError> value when no message is available. |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
4062979 | When Operators connected in a Directed Acyclic Graph (DAG) are executed in a multithreaded scheduler, it is not ensured that their execution order in the graph is adhered. |
4267272 | AJA drivers cannot be built with RDMA on IGX SW 1.0 DP iGPU due to missing nv-p2p.h . Expected to be addressed in IGX SW 1.0 GA. |
4384768 | No RDMA support on JetPack 6.0 DP and IGX SW 1.0 DP iGPU due to missing nv-p2p kernel module. Expected to be addressed in JP 6.0 GA and IGX SW 1.0 GA respectively. |
4190019 | Holoviz segfaults on multi-gpu setup when specifying device using the --gpus flag with docker run . Current workaround is to use CUDA_VISIBLE_DEVICES in the container instead. |
4210082 | v4l_camera example seg faults at exit. |
4339399 | High CPU usage observed with video_replayer_distributed application. While the high CPU usage associated with the GXF UCX extension has been fixed since v1.0, distributed applications using the MultiThreadScheduler (with the check_recession_period_ms parameter set to 0 by default) may still experience high CPU usage. Setting the HOLOSCAN_CHECK_RECESSION_PERIOD_MS environment variable to a value greater than 0 (e.g. 1.5 ) can help reduce CPU usage. However, this may result in increased latency for the application until the MultiThreadScheduler switches to an event-based multithreaded scheduler. |
4318442 | UCX cuda_ipc protocol doesn't work in Docker containers on x86_64. As a workaround, we are currently disabling the UCX cuda_ipc protocol on all platforms via the UCX_TLS environment variable. |
4325468 | The V4L2VideoCapture operator only supports YUYV and AB24 source pixel formats, and only outputs the RGBA GXF video format. Other source pixel formats compatible with V4L2 can be manually defined by the user, but they're assumed to be equivalent to RGBA8888. |
4325585 | Applications using MultiThreadScheduler may exit early due to timeouts. This occurs when the stop_on_deadlock_timeout parameter is improperly set to a value equal to or less than check_recession_period_ms , particularly if check_recession_period_ms is greater than zero. |
4301203 | HDMI IN fails in v4l2_camera on IGX Orin Devkit for some resolution or formats. Try the latest firmware as a partial fix. Driver-level fixes expected in IGX SW 1.0 GA. |
4384348 | UCX termination (either ctrl+c , press 'Esc' or clicking close button) is not smooth and can show multiple error messages. |
4481171 | Running the driver for a distributed applications on IGX Orin devkits fails when connected to other systems through eth1. A workaround is to use eth0 port to connect to other systems for distributed workloads. |
4458192 | In scenarios where distributed applications have both the driver and workers running on the same host, either within a Docker container or directly on the host, there's a possibility of encountering "Address already in use" errors. A potential solution is to assign a different port number to the HOLOSCAN_HEALTH_CHECK_PORT environment variable (default: 8777 ), for example, by using export HOLOSCAN_HEALTH_CHECK_PORT=8780 . |
Wayland: holoscan::viz::Init() with existing GLFW window fails. |
Holoscan SDK v2.1.0
Release Artifacts
- 🐋 Docker container: tag
v2.1.0-dgpu
andv2.1.0-igpu
- 🐍 Python wheel:
pip install holoscan==2.1.0
- 📦️ Debian packages:
2.1.0.1-1
- 📕 Documentation
See supported platforms for compatibility.
Release Notes
New Features and Improvements
Core
-
A report with execution time statistics for individual operators can now be enabled. This report will contain information like median, 90th percentile and maximum times for operator execution. Setting environment variable
HOLOSCAN_ENABLE_GXF_JOB_STATISTICS=true
enables this report (it is disabled by default as statistics collection may introduce a minor performance overhead). For more details see the documentation on the feature](https://docs.nvidia.com/holoscan/sdk-user-guide/gxf_job_statistics.html). -
The
holoscan.Tensor
object'sdata
property in the Python API now returns an integer (pointer address) instead of a NULL PyCapsule object, potentially avoiding confusion about data availability. Users can confirm the presence of data via the__array_interface__
or__cuda_array_interface__
properties. This change allows for direct access to the data pointer, facilitating debugging and performance optimization. -
The string representation of the
IOSpec
object, generated byIOSpec::to_yaml_node()
, includesConditionType
information in thetype
field. It correctly displayskNone
when no condition (ConditionType::kNone
in C++ andConditionType.NONE
in Python) is explicitly set.name: receiver io_type: kInput typeinfo_name: N8holoscan3gxf6EntityE connector_type: kDefault conditions: - type: kNone
-
Enhanced the macros (
HOLOSCAN_CONDITION_FORWARD_TEMPLATE
,HOLOSCAN_RESOURCE_FORWARD_TEMPLATE
,HOLOSCAN_OPERATOR_FORWARD_TEMPLATE
, etc.) by using the full namespace of the classes, improving their robustness and adaptability across different namespace scopes. -
Updated the
holoscan.core.py_object_to_arg()
method to allow conversion of Python objects toArg
objects usingYAML::Node
. This resolves type mismatches, such as when the underlying C++ parameter expects an int32_t type but Python uses int64_t. -
The Python OperatorSpec/ComponentSpec class exposes the inputs and outputs properties, providing direct access to the input and output IO specs. This enhancement simplifies the process of setting conditions on inputs and outputs.
def setup(self, spec: OperatorSpec): spec.input("data") # Set the NONE condition to the input port named `data`. spec.inputs["data"].condition(ConditionType.NONE) print(spec.inputs["data"])
-
Workflows where an operator connects to multiple downstream operators within the same fragment may see a minor performance boost. This is because of an internal refactoring in how connections between operators are made. Previously a GXF broadcast codelet was automatically inserted into the graph behind the scenes to broadcast the output to multiple receivers. As of this release, direct 1:N connection from an output port is made without the framework needing to insert this extra codelet to enable this.
-
fmt::format
support for printing theParameter
class has been added (there is no longer a need to call theget()
method to print out the contained value). This allows parameter values to be directly printed inHOLOSCAN_LOG_*
statements. For example :
MetaParameter p = MetaParameter<int>(5);
HOLOSCAN_LOG_INFO("Formatted parameter value: {}", p);
// can also pass parameter to fmt::format
std::string format_message = fmt::format("{}", p);
Operators/Resources
-
Most built-in operators now do additional validation of input tensors and will raise more helpful messages if the dimensions, data type or memory layout of the provided tensors is not as expected. Remaining operators (
InferenceOp
,InferenceProcessorOp
) will updated in the next release. -
BayerDemosaicOp
andFormatConverterOp
will now automatically perform host->device copies if needed for eithernvidia::gxf::VideoBuffer
orTensor
inputs. Previously these operators only did the transfer automatically fornvidia::gxf::VideoBuffer
, but not forTensor
and in the case ofFormatConverterOp
that transfer was only automatically done for pinned host memory. As of this release both operators will only copy unpinned system memory, leaving pinned host memory as-is. -
When creating Python bindings for C++ operators, it is now possible to register custom type conversion functions for user defined C++ types. These handle conversion to and from a corresponding Python type. See the newly expanded section on creating Python bindings for C++ operators for details.
-
As of this release, all provided Python operators support passing conditions such as
CountCondition
orPeriodicCondition
as positional arguments. In previous releases, there was a limitation that Python operators that wrapped an underlying C++ operator did not support this. As a concrete example, one could now pass aCountCondition
to limit the number of frames the visualization operator will run for.holoviz = HolovizOp( self, # add count condition to stop the application after short duration (i.e. for testing) CountCondition(self, count), name="holoviz", **self.kwargs("holoviz"), )
-
The AJA NTV2 dependency, and the corresponding AJA Source Operator, have been updated to use the latest official AJA NTV2 17.0.1 release. This new NTV2 version also introduces support for the KONA XM hardware.
-
The Holoviz operator now supports setting the camera for layers rendered in 3d (geometry layer with 3d primitives and depth map layer).
The camera eye, look at and up vectors can be initialized using parameters or dynamically changed at runtime by providing data at the respective input channels.
More information can be found in the documentation.
There is also a new C++ example holoviz_camera.cpp. -
The Holoviz operator now supports different types of camera pose outputs. In additional to the 4x4 row major projection matrix, a camera extrinsics model of type
nvidia::gxf::Pose3D
can now also be output. The output type is selected by setting thecamera_pose_output_type
parameter. -
The Holoviz operator now supports Wayland. Also the
run launch
command has been updated to support Wayland. -
The inference operator (
InferenceOp
) now supports a new optional parameter,temporal_map
, which can be used to specify a frame interval at which inference will be run. For example, setting a value of 10 for a given model will result in inference only being run on every 10th frame. Intermediate frames will output the result from the most recent frame at which inference was run. The interval value is specified per-model, allowing different inference models to be run at different rates. -
The existing asynchronous scheduling condition is now also available from Python (via
holoscan.conditions.AsynchronousCondition
). For an example of usage, see the new asynchronous ping example. -
We introduce the
GXFCodeletOp
andGXFComponentResource
classes, streamlining the import of GXF Codelets and Components into Holoscan applications. These additions simplify the setup process, allowing users to utilize custom GXF components more intuitively and efficiently.auto tx = make_operator<ops::GXFCodeletOp>( "tx", "nvidia::gxf::test::SendTensor", make_condition<CountCondition>(15), Arg("pool") = make_resource<GXFComponentResource>( "pool", "nvidia::gxf::BlockMemoryPool", Arg("storage_type") = static_cast<int32_t>(1), Arg("block_size") = 1024UL, Arg("num_blocks") = 2UL));
tx = GXFCodeletOp( self, "nvidia::gxf::test::SendTensor", CountCondition(self, 15), name="tx", pool=GXFComponentResource( self, "nvidia::gxf::BlockMemoryPool", name="pool", storage_type=1, block_size=1024, num_blocks=2, ), )
Please check out the examples in the examples/import_gxf_components directory for more information on how to use these new classes.
-
When calling
op_output.emit
from thecompute
method of a Python operator, it is now possible to provide a third argument that overrides the default choice of the type of object type emitted. This is sometimes needed to emit a certain C++ type from a native Python operator when connecting it to a different Python operator that wraps an underlying C++ operator. For example, one could emit a Python string as a C++std::string
instead of the Python string object viaop_output.emit(py_str, "port_name", "std::string")
. See additional examples and a table of the C++ types registered by default [here](https://docs.nvidia.com/holoscan/sdk-user-guide/holoscan_create_operator_...
Holoscan SDK v2.0.0
Release Artifacts
- 🐋 Docker container: tag
v2.0.0-dgpu
andv2.0.0-igpu
- 🐍 Python wheel:
pip install holoscan==2.0.0
- 📦️ Debian packages:
2.0.0.2-1
- 📕 Documentation
See supported platforms for compatibility.
Release Notes
New Features and Improvements
Core
make_condition
,make_fragment
,make_network_context
,make_operator
,make_resource
, and
make_scheduler
now accept a non-const
string or character array for thename
parameter.- A new event-based mult-thread scheduler (
EventBasedScheduler
) is available. It is an alternative to the existing, polling-basedMultiThreadScheduler
and can be used as a drop-in replacement. The only difference in parameters is that it does not takecheck_recession_period_ms
parameter, as there is no such polling interval for this scheduler. It should give similar performance to theMultiThreadScheduler
with a very short polling interval, but without the high CPU usage seen for the multi-thread scheduler in that case (due to constant polling for work by one thread). - When an exception is raised from the
Operator
methodsstart
,stop
orcompute
, that exception will first trigger the underlying GXF scheduler to terminate the application graph and then the exception will be raised by Holoscan SDK. This resolves an issue with inconsistent behavior from Python and C++ apps on how exceptions were handled and fixes a crash in C++ apps when an operator raised an exception from thestart
orstop
methods. - Now, when an exception occurs during the execution of a Holoscan application, it is propagated to
the application'srun
method, allowing users to catch and manage exceptions within their
application.
Previously, the Holoscan runtime would catch and log exceptions, with the application continuing
to run (in Python) or exit (in C++) without a clear indication of the exception's origin.
Users can catch and manage exceptions by enclosing therun
method in atry
block. - In the case of the
holoscan::Fragment::run_async
andholoscan.Application.run_async
methods
for C++ and Python, they returnstd::future
andconcurrent.futures.Future
respectively.
The revised documentation advises usingfuture.get()
in C++ andfuture.result()
in Python to
wait until the application has completed execution and to address any exceptions that occurred.
Operators
- V4L2 Video Capture: added support to set manual
exposure
andgain
values for cameras that support it. - Inference: one can now run multiple instances of the Inference operator in a single application without ressource conflicts.
Utils
- Can now build from source for iGPU (IGX iGPU, Jetpack) from a non-iGPU system (IGX dGPU, x86_64)
- The NGC container now supports packaging and running Holoscan Application Packages using the Holoscan CLI.
- CLI runner - better handling of the use of available GPUs by reading the package manifest file and check the system for available GPUs. New
--gpus
argument to override the default values.
Breaking Changes
-
The
VideoStreamRecorderOp
andVideoStreamReplayerOp
now work without requiring thelibgxf_stream_playback.so
extension. Now that the extension is unused, it has been removed from the SDK and should no longer be listed under theextensions
section of application YAML files using these operators. -
As of version 2.0, we have removed certain Python bindings to align with the unified logger interface:
- Removed APIs:
holoscan.logger.enable_backtrace()
holoscan.logger.disable_backtrace()
holoscan.logger.dump_backtrace()
holoscan.logger.should_backtrace()
holoscan.logger.flush()
holoscan.logger.flush_level()
holoscan.logger.flush_on()
- However, the following APIs remain accessible for Python. These are intended for logging in Holoscan's core or for C++ operators (e.g., using the
HOLOSCAN_LOG_INFO
macro), and are not designed for Python's logging framework. Python API users are advised to utilize the standardlogging
module for their logging needs:holoscan.logger.LogLevel
holoscan.logger.log_level()
holoscan.logger.set_log_level()
holoscan.logger.set_log_pattern()
- Removed APIs:
-
Several GXF headers have moved from
gxf/std
togxf/core
:parameter_parser.hpp
parameter_parser_std.hpp
parameter_registrar.hpp
parameter_storage.hpp
parameter_wrapper.hpp
resource_manager.hpp
resource_registrar.hpp
type_registry.hpp
-
Some C++ code for tensor interoperability has been upstreamed from Holoscan SDK into GXF. The public
holoscan::Tensor
class will remain, but there have been a small number of backward incompatible changes in related C++ classes and methods in this release. Most of these were used internally and are unlikely to affect existing applications.- supporting classes
holoscan::gxf::GXFTensor
andholoscan::gxf::GXFMemoryBuffer
have been removed. The DLPack functionality that was formerly inholoscan::gxf::GXFTensor
is now available upstream in GXF'snvidia::gxf::Tensor
. - The struct
holoscan::gxf::DLManagedTensorCtx
has been renamed toholoscan::gxf::DLManagedTensorContext
and is now just an alias fornvidia::gxf::DLManagedTensorContext
. It also has two additional fields (dl_shape
anddl_strides
to hold shape/stride information used by DLPack). holoscan::gxf::DLManagedMemoryBuffer
is now an alias tonvidia::gxf::DLManagedMemoryBuffer
- supporting classes
-
The GXF UCX extension, used in distributed applications, now sends data asynchronously by default, which can lead to issues such as insufficient memory on the transmitter side when a memory pool is used. Specifically, the concern is only for operators that have a memory pool and connect to an operator in a separate fragment of the distributed application. As a workaround, users can increase the
num_blocks
parameter to a higher value in theBlockMemoryPool
or use theUnboundedAllocator
to avoid the problem. This issue will be addressed in a future release by providing a more robust solution to handle the asynchronous data transmission feature of the UCX extension, eliminating the need for manual intervention (see Known Issue 4601414).-
For fragments using a
BlockMemoryPool
, thenum_blocks
parameter can be increased to a higher value to avoid the issue. For example, the following code snippet shows the existingBlockMemoryPool
resource being created with a higher number of blocks:recorder_format_converter = make_operator<ops::FormatConverterOp>( "recorder_format_converter", from_config("recorder_format_converter"), Arg("pool") = //make_resource<BlockMemoryPool>("pool", 1, source_block_size, source_num_blocks)); make_resource<BlockMemoryPool>("pool", 1, source_block_size, source_num_blocks * 2));
source_pool_kwargs = dict( storage_type=MemoryStorageType.DEVICE, block_size=source_block_size, #num_blocks=source_num_blocks, num_blocks=source_num_blocks * 2, ) recorder_format_converter = FormatConverterOp( self, name="recorder_format_converter", pool=BlockMemoryPool(self, name="pool", **source_pool_kwargs), **self.kwargs("recorder_format_converter"), ) )
-
Since the underlying UCXTransmitter attempts to send the emitted data regardless of the status of the downstream Operator input port's message queue, simply doubling the
num_blocks
may not suffice in cases where the receiver operator's processing time is slower than that of the sender operator. -
If you encounter the issue, consider using the
UnboundedAllocator
instead of theBlockMemoryPool
to avoid the problem. TheUnboundedAllocator
does not have a fixed number of blocks and can allocate memory as needed, though it can cause some overhead due to the lack of a fixed memory pool size and may lead to memory exhaustion if the memory is not released in a timely manner.
The following code snippet shows how to use theUnboundedAllocator
:... Arg("pool") = make_resource<UnboundedAllocator>("pool");
from holoscan.resources import UnboundedAllocator ... pool=UnboundedAllocator(self, name="pool"), ...
-
Bug fixes
Issue | Description |
---|---|
4381269 | Fixed a bug that caused memory exhaustion when compiling the SDK in the VSCode Dev Container (using 'Tasks: Run Build Task') due to the missing CMAKE_BUILD_PARALLEL_LEVEL environment variable. Users can specify the number of jobs with the --parallel option (e.g., ./run vscode --parallel 16 ). |
4569102 | Fixed an issue where the log level was not updated from the environment variable when multiple Application classes were created during the session. Now, the log level setting in Application class allows for a reset from the environment variable if overridden. |
4578099 | Fixed a segfault in FormatConverterOp if used with a BlockMemoryPool with insufficient capacity to create the output tensor. |
4571581 | Fixed an issue where the documentation for the built-in operators was either missing or incorrectly rendered. |
4591763 | Application crashes if an exception is thrown from Operator::start or Operator::stop |
4595680 | Fixed an issue that caused the Inference operator to fail when multiple instances were composed in a... |
Holoscan SDK v1.0.3
Release Artifacts
- 🐋 Docker container: tag
v1.0.3-dgpu
andv1.0.3-igpu
- 🐍 Python wheels:
holoscan==1.0.3
- 📦️ Debian packages:
1.0.3.2-1
(from the cuda repository) - 📕 Documentation
Release Notes
New Features and Improvements
Core
- Allow operator input and output ports to have matching names
- Application graphs with cycles are now supported
(example) - Cycles in the graph are also supported in Data Flow Tracking
- An informative error message is now raised if an unsupported condition type is provided to
IOSpec::condition
. - User-defined operators can now define parameters that are of type
complex<float>
orcomplex<double>
. These parameters can either be parsed from a YAML config (e.g. using a string like "1.0 + 2.0j") or passed as aholoscan::Arg
to the operator constructor. - Holoscan tensors containing data of type
complex<float>
orcomplex<double>
can now be used. - Python applications can now send CuPy, NumPy or other tensor types with complex-valued data between fragments of a multi-fragment application. Previously, this only worked within a single fragment.
- Many C++ API
description
methods and corresponding Python API__repr__
methods have been improved.- The
IOSpec
class now has adescription
method and corresponding Python__repr__
method. - A bug was fixed where the
Arg
class__repr__
could raiseUnicodeDecodeError
for uint8_t or int8_t argument types - The
NetworkContext
andScheduler
print more comprehensive information. - Python bindings for GXF conditions, resources and operators have an improved
__repr__
that makes use of the underlying C++ description methods.
- The
- The
HOLOSCAN_UCX_PORTS
environment variable allows users to define preferred port numbers for the SDK's inter-node communication in a distributed application, especially in environments where specific ports need to be predetermined, such as Kubernetes. - A
Condition
orResource
class can be added to a Python operator after construction via itsadd_arg
method. - Distributed applications can now leverage RDMA transports with MLNX_OFED drivers. Tested with RoCE.
- The
HOLOSCAN_HEALTH_CHECK_PORT
environment variable allows users to define a port number for the SDK's health check endpoint in a distributed application. - A set of available keys in an
Application
orFragments
's YAML configuration file can now be determined via a newconfig_keys()
method in the C++ API orconfig_keys
method from Python. - Debugging (tracing and profiling) of Python operators is now fully supported.
- Previously, the
compute
,initialize
,start
, andstop
methods of the Holoscan Operator were not compatible with Python tracing/profiling in earlier releases. - Debugging methods of Python operators with the VSCode/PyCharm debugger using PyDev.Debugger (pydevd) is now feasible, as well as profiling or gathering coverage data using cProfile or coverage.py.
- For comprehensive information, refer to the Debugging section in the SDK User Guide.
- Previously, the
Operators
- HoloViz
- Inference
- The ONNX Runtime (ORT) inference backend is now a plugin, like the Torch backend, allowing you to use the inference operator without requiring an installation of ORT when using other backends (like TensorRT or Torch).
Utils
- Added a Dockerfile that contains only runtime dependencies. This Dockerfile can be built by running
./run build_run_image
at the top of the repository, creating an image that is ~8.6 GB vs. the ~13 GB build container from./run build_image
. (doc) - The
run
script in the git repository had a couple of updates and improvements, including:- Allow building as root
- Allow running of the build container without display
- Naming build image, build directories, and install directories with the target architecture and GPU (ex:
build
->build-aarch64-dgpu
) - Support building on system without tty support
- Support running on system without xhost support
- Added more flags (see
./run help
and./run <cmd> --help
for details)
Packaging
- Mellanox OFED user libraries were added to the NGC container to allow the use of RDMA transports from the container.
Documentation
- The user guide source code and tooling is now released on GitHub (link)
Breaking Changes
- H264 operator and applications were moved from the SDK to HoloHub (MR)
- For distributed applications, there is a change to the emit/receive behavior for array-like objects (e.g. PyTorch tensor) between operators within a fragment. Previously (in v0.6.x), the array-like object type was always preserved for within-fragment emit/receive. However, now now any host array-like will be recevied as a NumPy array (and any device array-like will be received as a CuPy array). Making within-fragment emit/receive behavior consistent with between-fragment emit/receive behavior was necessary to implement the fix for issue 4290043.
- Building against Ubuntu 22.04, debian packages and python wheels require
GLIBC_2.35
or above.
Bug fixes
Issue | Description |
---|---|
4185976 | Cycle in a graph is not supported. As a consequence, the endoscopy tool tracking example using input from an AJA video card in enabled overlay configuration is unfunctional. This is planned to be addressed in the next version of the SDK. |
4196152 | Getting "Unable to find component from the name ''" error message when using InferenceOp with Data Flow Tracking enabled. |
4211747 | Communication of GPU tensors between fragments in a distributed application can only use device 0 |
4212743 | Holoscan CLI packager copies into the App Package the unrelated files and folders in the same folder than the model file. |
4232453 | A segfault occurs if a native Python operator __init__ assigns a new attribute that overrides an existing base class attribute or method. A segfault will also occur if any exception is raised during Operator.__init__ or Application.__init__ before the parent class __init__ was called. |
4206197 | Distributed apps hang if multiple input/output ports are connected between two operators in different fragments. |
3599303 | Linux kernel is not built with security hardening flags. Future releases will include a Linux kernel built with security hardening flags. |
4187787 | TensorRT backend in the Inference operator prints Unknown embedded device detected. Using 52000MiB as the allocation cap for memory on embedded devices on IGX Orin (iGPU). Addressed in TensorRT 8.6+. |
4194109 | AppDriver is executing fragments' compose() method which can be avoided. |
4260969 | App add_flow causes issue if called more than once between a pair of operators. |
4265393 | Release 1.0-ea1 and 1.0-ea2 fail to run distributed applications with workers on two or more nodes. |
4272363 | A segfault may occur if an operator's output port containing GXF Tensor data is linked to multiple operators within the MultiThreadScheduler. |
4290043 | Bug in Python implicit broadcast of non-TensorMap types when at least one target operator is in a different fragment. |
4293729 | Python application using MultiThreadScheduler (including distributed application) may fail with GIL related error if SDK was compiled in debug mode. |
4101714 | Vulkan applications fail (vk::UnknownError ) in containers on iGPU due to missing iGPU device node being mounted in the container. Workaround documented in run instructions. |
3881725 | VK_ERROR_INITIALIZATION_FAILED with segmentation fault while running High-Speed Endoscopy gxf/cpp app on Clara AGX developer kits. Fix available in CUDA drivers 520. Workaround implemented since v0.4 to retry automatically. |
4293741 | Python application with more than two operators (mixed use of pure Python operator and operator wrapping C++ operator), using MultiThreadScheduler (including distributed app) and sending Python tensor can deadlock at runtime. |
4313690 | Failure to initialize BayerDemosaicOp in applications using the C++ API |
4187826 | Torch backend in the Inference operator is not supported on Tegra's integrated GPU. |
4336947 | The dev_id parameter of the CudaStreamPool resource is ignored. |
4344061 | Native Python operator overrides of the start, stop or initialize methods don't handle exceptions properly |
4344408 | The distributed application displays an error message if port 8777 is already in use. |
4363945 | Checking if a key exists in an application's config file results in an error being logged. |
Fixed bad cast exception when defining optional ports enablement (buffer input, output, camera pose) for the Holoviz operator from a YAML configuration file. | |
Fixed invalid stride alignment of... |
Holoscan SDK v0.6.0
Release Artifacts
- 🐋 Docker container: tag
v0.6.0-dgpu
andv0.6.0-igpu
- 🐍 Python wheel:
holoscan==0.6.0
- 📦️ Debian package: tags
v0.6.0-amd64
andv0.6.0-arm64
- 📕 Documentation
Release Notes
New Features and Improvements
Core
- Multi-fragments application support for distributed workloads (doc)
- Async run (doc)
- Multithread scheduler (doc1, doc2)
- Async and Periodic conditions (doc1, doc2)
- Realtime and Manual Clock classes (doc)
- Optional GXF Parameter support
- Topologically sorted Graph initialization (doc)
- Data Flow Tracking (doc)
- Lazy import of individual operator modules in python
Operators
- A V4L2 operator supporting input video streams USB and HDMI IN was added (example)
- Improvements to the Inference (previously MultiAIInference) operator (doc), including:
- Torch support
- Multi backend support
- Multi I/O support
- Improvements to the Holoviz (visualization) operator (doc), including:
- Dynamic text support
- Multi-views support
- Multi-GPU data transfer support
Utils
- Application packager and runner (doc)
- A new
HOLOSCAN_LOG_FORMAT
environment variable has been added to allow user to modify logger message format at runtime - Auto loading of log verbosity environment variable (
HOLOSCAN_LOG_LEVEL
) and YAML config path (HOLOSCAN_CONFIG_PATH
) - Script to decode GXF entities files (doc)
HoloHub
- Volume rendering support with ClaraViz (operator, application)
- Development container support (doc)
Documentation
- New Getting Started section
- Improved and consolidated SDK Installation section
- Various additions and improvements
Breaking API Changes
Core
- The function
holoscan::InputContext::receive
has been modified to returnholoscan::expected<DataT, holoscan::RuntimeError>
instead ofstd::shared_ptr<DataT>
where it returns either a valid value or an error (with the type and explanation of the error). Note that IO objects are not all assumed to be wrapped in astd::shared_ptr
anymore. - Messages of type
gxf::Entity
between GXF based Operators and Holoscan Native Operators have been modified to typeholoscan::TensorMap
in C++ anddict
type of objects in Python .
Operators
- The deprecated TensorRT inference operator was removed in favor of the Multi AI Inference operator, which was renamed to Inference operator (doc):
- Include headers:
holoscan/operators/tensor_rt/tensor_rt_inference.hpp
removedholoscan/operators/multiai_inference/multiai_inference.hpp
renamed toholoscan/operators/inference/inference.hpp
holoscan/operators/multiai_postprocessor/multiai_postprocessor.hpp
renamed toholoscan/operators/inference_processor/inference_processor.hpp
- C++ classes:
holoscan::ops::TensorRtInferenceOp
removedholoscan::ops::MultiAIInferenceOp
renamed toholoscan::ops::InferenceOp
holoscan::ops::MultiAIPostprocessorOp
renamed toholoscan::ops::InferenceProcessorOp
- CMake targets:
holoscan::ops::tensor_rt
removedholoscan::ops::multiai_inference
renamed toholoscan::ops::inference
holoscan::ops::multiai_postprocessor
renamed toholoscan::ops::inference_processor
- Include headers:
Utils
- The function
holoscan::load_env_log_level
has been removed. TheHOLOSCAN_LOG_LEVEL
environment is now loaded automatically.
HoloHub
- The class
ops::VideoDecoderOp
has been replaced with the classesops::VideoDecoderRequestOp
,ops::VideoDecoderResponseOp
andops::VideoDecoderContext
- The class
ops::VideoEncoderOp
has been replaced with the classesops::VideoEncoderRequestOp
,ops::VideoEncoderResponseOp
andops::VideoEncoderContext
Bug fixes
Issue | Description |
---|---|
3762996 | nvidia-peermem.ko fails to load using insmod on Holoscan devkits in dGPU mode. |
4048062 | Warning or error when deleting TensorRT runtime ahead of deserialized engines for some versions of TensorRT |
4036186 | H264 encoder/decoder are not supported on iGPU |
Supported Platforms
Note: This release is intended for use with the listed platforms only. NVIDIA does not provide support for this release on products other than those listed below.
Platform | OS |
---|---|
NVIDIA IGX Orin Developer Kit | NVIDIA HoloPack 2.0 (L4T r35.3.1) or Meta Tegra Holoscan 0.6.0 (L4T r35.3.1) |
NVIDIA Jetson AGX Orin Developer Kit | NVIDIA JetPack r35.1.1 |
NVIDIA Clara AGX Developer Kit | NVIDIA HoloPack 1.2 (L4T r34.1.2) or Meta Tegra Holoscan 0.6.0 (L4T r35.3.1) |
x86_64 platforms with Ampere GPU or above(tested with RTX6000 and A6000) |
Ubuntu 20.04 |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
3878494 | Inference fails after tensorrt engine file is first created using BlockMemoryPool . Fix available in TensorRT 8.4.1. Use UnboundedAllocator as a workaround. |
3599303 | Linux kernel is not built with security hardening flags. Future releases will include a Linux kernel built with security hardening flags. |
3881725 | VK_ERROR_INITIALIZATION_FAILED with segmentation fault while running High-Speed Endoscopy gxf/cpp app on Clara AGX developer kits. Fix available in CUDA drivers 520. Workaround implemented since v0.4 to retry automatically. |
4047688 | H264 applications are missing dependencies (nvidia-l4t-multimedia-utils ) to run in the arm64 dGPU container |
4062979 | When Operators connected in a Directed Acyclic Graph (DAG) are executed in a multithreaded scheduler, it is not ensured that their execution order in the graph is adhered. |
4068454 | Crash on systems with NVIDIA and non-NVIDIA GPUs. Workaround documented in Troubleshooting section of the GitHub README. |
4101714 | Vulkan applications fail (vk::UnknownError ) in containers on iGPU due to missing iGPU device node being mounted in the container. Workaround documented in run instructions. |
4171337 | AJA with RDMA is not working on integrated GPU (IGX or AGX Orin) due to conflicts between the nvidia-p2p and nvidia driver symbols (nvidia_p2p_dma_map_pages ). Fixed in JetPack 5.1.2, expected in HoloPack 2.1 |
4185260 | H264 application process hangs after X11 video exit. |
4185976 | Cycle in a graph is not supported. As a consequence, the endoscopy tool tracking example using input from an AJA video card in enabled overlay configuration is unfunctional. This is planned to be addressed in the next version of the SDK. |
4187826 | Torch backend in the Inference operator is not supported on Tegra's integrated GPU. |
4187787 | TensorRT backend in the Inference operator prints Unknown embedded device detected. Using 52000MiB as the allocation cap for memory on embedded devices on IGX Orin (iGPU). Addressed in TensorRT 8.6+. |
4190019 | Holoviz segfaults on multi-gpu setup when specifying device using the --gpus flag with docker run . Current workaround is to use CUDA_VISIBLE_DEVICES in the container instead. |
4196152 | Getting "Unable to find component from the name ''" error message when using InferenceOp with Data Flow Tracking enabled. |
4199282 | H264 applications may fail on x86_64 du... |
Holoscan SDK v0.5.1
Release Artifacts
- 🐋 Docker container: tag
v0.5.1-dgpu
andv0.5.1-igpu
- 🐍 Python wheel:
holoscan==0.5.1
- 📦️ Debian package: tags
v0.5.1-amd64
andv0.5.1-arm64
- 📕 Documentation
Release Notes
Additions
This release of the Holoscan SDK provide the following additions:
Support for the NVIDIA IGX Orin Developer Kit
The Holoscan SDK 0.5.1 adds support for the NVIDIA IGX Orin Developer Kit in both iGPU or dGPU modes. That support is enabled by the release of the Holopack 2.0 Developer Preview now available through the latest version of the SDK Manager. Python wheels and Debian packages for the arm64/aarch64 architecture support both iGPU and dGPU. Starting with 0.5.1, the Holoscan container on NGC now offers two separate tags, one for iGPU and one for dGPU.
Lazy loading of python modules
Python users might be interested in using the Holoscan SDK without requiring to import every operators, which requires to have all their dependencies available in the python environment (example: TensorRT...). With 0.5.1, a user can create an Holoscan application by only importing the modules they require, for example, importing holoscan.core
only and not holoscan.operators
.
Note: in 0.6.0, the operators will be broken down in separate modules to offer further importing granularity and lower dependency requirements.
ffmpeg support in the Holoscan container on NGC
Facilitates conversions of standard video formats to gxf entities in the Holoscan containers (example)
Improved L4T Compute Assist container
The L4T Compute Assist container - used to run compute workloads on the iGPU of a developer kit configured for dGPU - now includes the deviceQuery
executable to facilitate validating its configuration, along with updated troubleshooting steps.
Issues Fixed
Issue | Description |
---|---|
- | Fixed instructions in documentation for datasets download and debian package installation |
Supported Platforms
Note: This release is intended for use with the listed platforms only. NVIDIA does not provide support for this release on products other than those listed below.
Platform | OS |
---|---|
NVIDIA Clara AGX Developer Kit | - NVIDIA Holopack 1.2 (L4T r34.1.2) - Meta Tegra Holoscan 0.5.0 (L4T r35.2.1) |
NVIDIA IGX Orin [ES] Developer Kit | - NVIDIA Holopack 1.2 (L4T r34.1.2) - Meta Tegra Holoscan 0.5.0 (L4T r35.2.1) |
NVIDIA IGX Orin Developer Kit | - NVIDIA Holopack 2.0 (L4T r35.4.0) - Meta Tegra Holoscan 0.5.1 (L4T r35.3.1) |
x86_64 platforms with Ampere GPU or above(tested with RTX6000 and A6000) |
Ubuntu 20.04 |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
3878494 | Inference fails after tensorrt engine file is first created using BlockMemoryPool . Fix available in TensorRT 8.4.1. Use UnboundedAllocator as a workaround. |
3762996 | nvidia-peermem.ko fails to load using insmod on Holoscan devkits in dGPU mode. Install nvidia-peer-memory following the RDMA instructions in the Holoscan SDK User Guide. |
3655489 | Installing dGPU drivers can remove nvgpuswitch.py script from the executable search path. Explicitly including /opt/nvidia/l4t-gputools/bin in the PATH environment variable ensures this script can be found for execution. |
3599303 | Linux kernel is not built with security hardening flags. Future releases will include a Linux kernel built with security hardening flags. |
3633688 | RDMA on the NVIDIA IGX Orin [ES] Developer Kit (holoscan-devkit) is not functional. PCIe switch firmware update fixed the issue. RDMA for the Clara AGX Developer Kit is functional and unaffected by this issue. |
3881725 | VK_ERROR_INITIALIZATION_FAILED with segmentation fault while running High-Speed Endoscopy gxf/cpp app. Fix available in CUDA drivers 520. Workaround implemented in v0.4 to retry automatically. |
4048062 | Warning or error when deleting TensorRT runtime ahead of deserialized engines for some versions of TensorRT |
4036186 | H264 encoder/decoder are not supported on iGPU |
4047688 | H264 applications are not able to run in the arm64 dGPU container |
4101714 | --privileged permission required to run rendering applications from the Holoscan iGPU container on IGX Orin Developer Kit with Holopack 2.0 DP |
4116861 | H264 video encoding fails on IGX Orin Developer Kit with Holopack 2.0 DP |
Holoscan SDK v0.5.0
Release Artifacts
- 🐋 Docker container: tag
v0.5.0
- 🐍 Python wheel:
holoscan==0.5.0
- 📦️ Debian package: tags
v0.5.0-amd64
andv0.5.0-arm64
- 📕 Documentation
Release Notes
New Features
This release of the Holoscan SDK along with additions to HoloHub provide the following main features:
H264 encoder/decoder support
Operators to support H264 bitstream accelerated encoder and decoder were added to HoloHub, as illustrated by two new applications: h264_video_decode and h264_endoscopy_tool_tracking.
iGPU compute support on Holoscan Developer Kits
The L4T Compute Assist container is now available on NGC to perform computation on the integrated GPU (iGPU) of Holoscan Developer Kits configured to use their discrete GPU (dGPU), allowing to run workloads on both GPUs in parallel.
Use Holoscan operators in GXF applications
Infrastructure and documentation were added to wrap Holoscan operators as GXF codelets so they can be used by other frameworks which use GXF extensions.
x86_64 physical I/O support
The Holoscan SDK now officially support physical I/O on x86_64 platforms. The High Speed endoscopy application on HoloHub has been tested with Rivermax/GPU Direct RDMA support and offers similar performances as previously reported with the Holoscan Developer Kits.
Depth-map rendering
The Holoscan SDK visualization module (referred to as Holoviz) adds depth-map rendering capabilities to support displaying inference results with depth information.
New examples
The Holoscan SDK now provides a new suite of examples with associated step-by-step documentation to better introduce users to the SDK, taking them from a Hello World example to an application that deploys an ultrasound segmentation inference. Additional examples are also available to demonstrate how to integrate sensors and third-party frameworks into their workflow.
Changes from previous release
- All sample applications along with domain specific operators were migrated from the Holoscan SDK
to HoloHub. - Most operators have been transitioned to native implementations.
Issues Fixed
Issue | Description |
---|---|
3834424 | Ultrasound segmentation application is not functional on NVIDIA IGX Orin [ES] Developer Kit with iGPU configuration in deployment stack |
3842899 | High-Speed Endoscopy application is not supported in deployment stack |
3897810 | Applications not working on x86_64 systems with multiple GPUs |
3936290 | Cannot run exclusive display from docker container |
Supported Platforms
Note: This release is intended for use with the listed platforms only. NVIDIA does not provide support for this release on products other than those listed below.
Platform | OS |
---|---|
NVIDIA Clara AGX Developer Kit | - NVIDIA Holopack 1.2 (L4T r34.1.2) - Meta Tegra Holoscan 0.5.0 (L4T r35.2.1) |
NVIDIA IGX Orin [ES] Developer Kit | - NVIDIA Holopack 1.2 (L4T r34.1.2) - Meta Tegra Holoscan 0.5.0 (L4T r35.2.1) |
NVIDIA IGX Orin Developer Kit | Meta Tegra Holoscan 0.5.0 (L4T r35.2.1) |
x86_64 platforms with Ampere GPU or above(tested with RTX6000 and A6000) |
Ubuntu 20.04 |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
3878494 | Inference fails after tensorrt engine file is first created using BlockMemoryPool . Fix available in TensorRT 8.4.1. Use UnboundedAllocator as a workaround. |
3762996 | nvidia-peermem.ko fails to load using insmod on Holoscan devkits in dGPU mode. Install nvidia-peer-memory following the RDMA instructions in the Holoscan SDK User Guide. |
3655489 | Installing dGPU drivers can remove nvgpuswitch.py script from the executable search path. Explicitly including /opt/nvidia/l4t-gputools/bin in the PATH environment variable ensures this script can be found for execution. |
3599303 | Linux kernel is not built with security hardening flags. Future releases will include a Linux kernel built with security hardening flags. |
3633688 | RDMA on the NVIDIA IGX Orin [ES] Developer Kit (holoscan-devkit) is not functional. PCIe switch firmware update fixed the issue. RDMA for the Clara AGX Developer Kit is functional and unaffected by this issue. |
3881725 | VK_ERROR_INITIALIZATION_FAILED with segmentation fault while running High-Speed Endoscopy gxf/cpp app. Fix available in CUDA drivers 520. Workaround implemented in v0.4 to retry automatically. |
4048062 | Warning or error when deleting TensorRT runtime ahead of deserialized engines for some versions of TensorRT |
4036186 | H264 encoder/decoder are not supported on iGPU |
4047688 | H264 applications are missing dependencies (nvidia-l4t-multimedia-utils ) to run in the arm64 dGPU container |
Holoscan SDK v0.4.1
Release Notes
Changes from previous release
- HoloPack: update to version 1.2
- Python: throw warnings instead of exceptions if a GXF extension cannot be loaded, to unblock execution if the operator using that extension is not needed by the application.
Issues Fixed
- NGC container: fix issue related to expired signing key preventing to call
apt update
- Source: fix issues in Dockerfile related to missing pinned dependency and expired signing key
Please see the list of known issues below for more information.
Supported Platforms
Note: This release is intended for use with the listed platforms only. NVIDIA does not provide support for this release on products other than those listed below.
Description | Supported Version |
---|---|
Supported NVIDIA® Tegra® Linux Driver Package (L4T) | NVIDIA® Holopack 1.2 -- R34.1.2 |
Supported Jetson Platforms | Holoscan Developer Kits |
Supported x86_64 Platforms | Ubuntu 20.04 with Ampere GPU or above (tested with RTX6000 and A6000) |
Supported Software for Clara AGX Developer Kit with NVIDIA® RTX6000 and IGX Orin Developer Kit with NVIDIA® A6000 |
NVIDIA® Driver 510.73.08 CUDA 11.6.1 TensorRT 8.2.3 GXF 2.5 AJA NTV2 SDK 16.2 |
Known Issues
This section supplies details about issues discovered during development and QA but not resolved in this release.
Issue | Description |
---|---|
3878494 | Inference fails after tensorrt engine file is first created using BlockMemoryPool |
3762996 | nvidia-peermem.ko fails to load using insmod on Holoscan devkits in dGPU mode. Install nvidia-peer-memory following the RDMA instructions in the Holoscan SDK User Guide. |
3655489 | Installing dGPU drivers can remove nvgpuswitch.py script from the executable search path. Explicitly including /opt/nvidia/l4t-gputools/bin in the PATH environment variable ensures this script can be found for execution. |
3599303 | Linux kernel is not built with security hardening flags. Future releases will include a Linux kernel built with security hardening flags. |
3633688 | RDMA on the NVIDIA IGX Orin Developer Kit (holoscan-devkit) is not functional. PCIe switch firmware update fixed the issue. RDMA for the Clara AGX Developer Kit is functional and unaffected by this issue. |
3834424 | Ultrasound segmentation application is not functional on NVIDIA IGX Orin Developer Kit (holoscan-devkit) with iGPU configuration in deployment stack |
3842899 | High-Speed Endoscopy application is not supported in deployment stack. |
3881725 | VK_ERROR_INITIALIZATION_FAILED with segmentation fault while running High-Speed Endoscopy gxf/cpp app (workaround implemented in v0.4 and fix in available in 520 drivers) |
3897810 | Applications not working on x86_64 systems with multiple GPUs |
3936290 | Cannot run exclusive display from docker container |