Plug-ins

VVAS is based on the GStreamer framework. The two types of VVAS GStreamer plug-ins are custom plug-ins and infrastructure plug-ins. This section describes the VVAS GStreamer plug-ins, their input, outputs, and control parameters. The plug-ins source code is available in the vvas-gst-plugins folder of the VVAS source tree. This section covers the plug-in that are common for Edge as well as cloud solutions. There are few plug-ins that are specific to Edge/Embedded platforms and are covered in Plugins for Embedded platforms. Similarly there are few plug-ins that are specific to Cloud/Data Center platforms and these are covered in Plugins for Data Center Platform. The following table lists the VVAS GStreamer plug-ins.

Table 1: GStreamer Plug-ins

Plug-in Name	Functionality
vvas_xmetaaffixer	Plug-in to scale and attach metadata to the frame of different resolutions.
vvas_xabrscaler	Hardware accelerated scaling and color space conversion.
vvas_xmultisrc	A generic infrastructure plug-in: 1 input, N output, supporting transform processing.
vvas_xfilter	A generic infrastructure plug-in: 1 input, 1 output, supporting pass-through, in-place, and transform processing.
vvas_xinfer	An inference plug-in using vvas_xdpuinfer acceleration software library and attaches inference output as GstInferenceMeta to input buffer. Also, this plug-in does optinal pre-processing required for inference using vvas_xpreprocess acceleration software library.

Custom Plug-ins

There are specific functions, like video decoder, encoder, and meta affixer where the requirements are difficult to implement in an optimized way using highly simplified and generic infrastructure plug-ins framework. Hence, these functions are implemented using custom GStreamer plug-ins. This section covers details about the custom plug-ins.

vvas_xmetaaffixer

The metaaffixer plug-in, vvas_xmetaaffixer, is used to scale the incoming metadata information for the different resolutions. A machine learning (ML) operation can be performed on a different frame resolution and color format than the original frame, but the metadata might be associated with the full resolution, original frame. The vvas_metaaffixer has two types of pads, master pads and slave pads. Input pads are pads on request; the number of input pads can be created based on the number of input sources. There is one mandatory master input pad (sink pad) that receives the original/reference metadata. Other input pads are referred to as slave pads. Metadata received on the master sink pad is scaled in relation to the resolution of each of the slave sink pads. The scaled metadata is attached to the buffer going out of the output (source) slave pads. There can be up to 16 slave pads created as required. For implementation details, refer to vvas_xmetaaffixer source code

Input and Output

This plug-in is format agnostic and can accept any input format. It operates only on the metadata. The vvas_xmetaaffixer plug-in supports the GstInferenceMeta data structure. For details about this structure, refer to the VVAS Inference Metadata section.

Control Parameters and Plug-in Properties

Table 2: vvas_xmetaaffixer Plug-in Properties

Property Name	Type	Range	Default	Description
sync	Boolean	True/False	True	This property is to enable the synchronization between master and slave pads buffers. If sync=true is set, then the metadata is scaled and attached to buffers on slave pads that have matching PTS or PTS falls within frame duration of the buffer on the master sink pad. If sync=false is set on the element, then the metadata is scaled and attached to all the buffers on the slave pads. If this option is used, there is possibility that the metadata is not suitable for the frames/buffers that are not corresponding to the frames/buffers on the master pad.
timeout	Int64	-1 to 9223372036854	2000	Timeout in millisec.

Pad naming syntax

The pad naming syntax is listed, and the following image shows the syntax:

• MetaAffixer master input pad should be named sink_master.

• MetaAffixer master output pad should be named src_master.

• MetaAffixer slave input pad should be named sink_slave_0, sink_slave_1.

• MetaAffixer slave output pad should be named src_slave_0, src_slave_1, src_slave_2.

Example Pipelines

This section covers the example pipelines using the metaaffixer plug-in.

gst-launch-1.0 videotestsrc num-buffers=1 \
! video/x-raw, width=1920, height=1080, format=NV12 \
! queue \
! videoconvert \
! queue \
! ima.sink_master vvas_xmetaaffixer name=ima ima.src_master \
! queue \
! fakesink videotestsrc num-buffers=1 \
! video/x-raw, width=1920, height=1080, format=NV12 \
! queue \
! videoconvert \
! video/x-raw, width=1920, height=1080, format=YUY2 \
! ima.sink_slave_0 ima.src_slave_0 \
! queue \
! fakesink -v

vvas_xabrscaler

In adaptive bit rate (ABR) use cases, one video is encoded at different bit rates so that it can be streamed in different network bandwidth conditions without any artifacts. To achieve this, input frame is decoded, resized to different resolutions and then re-encoded. vvas_xabrscaler is a plug-in that takes one input frame and can produce several outputs frames having different resolutions and color formats. The vvas_xabrscaler is a GStreamer plug- in developed to accelerate the resize and color space conversion functionality. For more implementation details, refer to vvas_xabrscaler source code.

This plug-in supports:

• Single input, multiple output pads

• Color space conversion

• Resize

• Each output pad has independent resolution and color space conversion capability.

Important

The vvas_xabrscaler plug-in controls the multiscaler kernel. If your application uses this plug-in, then make sure that multi-scaler kernel is included in your hardware design.

Important

Make sure that the multi-scaler hardware kernel supports maximum resolution required by your application.

As a reference, maximum resolution supported by multi-scaler kernel in Smart Model Select example design can be found in multi-scaler kernel config

Prerequisite

This plug-in requires the multiscaler kernel to be available in the hardware design. See Multiscaler Kernel

Input and Output

This plug-in accepts buffers with the following color format standards:

• RGBx

• YUY2

• r210

• Y410

• NV16

• NV12

• RGB

• v308

• I422_10LE

• GRAY8

• NV12_10LE32

• BGRx

• GRAY10_LE32

• BGRx

• UYVY

• BGR

• RGBA

• BGRA

• I420

Important

Make sure that the color formats needed for your application are supported by the multi-scaler hardware kernel.

As a reference, multi-scaler configuration for smart model select example design can be found in multi-scaler configuration

Control Parameters and Plug-in Properties

The following table lists the GStreamer plug-in properties supported by the vvas_xabrscaler plug-in.

Table 3: vvas_xabrscaler Plug-in Properties

Property Name	Type	Range	Default	Description
avoid-output-copy	Boolean	true/false	False	Avoid output frames copy on all source pads even when downstream does not support GstVideoMeta metadata
enable-pipeline	Boolean	true/false	false	Enable buffer pipelining to improve performance in non zero-copy use cases
in-mem-bank	Unsigned int	0 - 65535	0	VVAS input memory bank to allocate memory
out-mem-bank	Unsigned int	0 - 65535	0	VVAS o/p memory bank to allocate memory
xclbin-location	string	N/A	./binary_container_1 xclbin	The location of xclbin.
kernel-name	string	N/A	v_multi_scaler: multi_scaler_1	Kernel name and instance separated by a colon.
dev-idx	integer	0 to 31	0	Device index This is valid only in PCIe/ Data Center platforms.
ppc	integer	1, 2, 4	2	Pixel per clock supported by a multi- scaler kernel
scale-mode	integer	0, 1, 2	0	Scale algorithm to use: 0:BILINEAR 1:BICUBIC 2:POLYPHASE
coef-load-type	integer	0 => Fixed 1 => Auto	1	Type of filter Coefficients to be used: Fixed or Auto generated
num-taps	integer	6=>6 taps 8=>8 taps 10=>10 taps 12=>12 taps	1	Number of filter taps to be used for scaling
alpha-b	float	0 to 128	0	Mean subreaction for blue channel
alpha-g	float	0 to 128	0	Mean subreaction for green channel
alpha-r	float	0 to 128	0	Mean subreaction for red channel
beta-b	float	0 to 1	1	Scaling for blue channel
beta-g	float	0 to 1	1	scaling for green channel
beta-r	float	0 to 1	1	scaling for red channel

Example Pipelines

One input one output

The following example configures vvas_xabrscaler in one input and one output mode. The input to the scaler is 1280 x 720, NV12 frames that are resized to 640 x 360 resolution, and the color format is changed from NV12 to BGR.

gst-launch-1.0 videotestsrc num-buffers=100 \
! "video/x-raw, width=1280, height=720, format=NV12" \
! vvas_xabrscaler xclbin-location="/usr/lib/dpu.xclbin" kernel-name=v_multi_scaler:v_multi_scaler_1 \
! "video/x-raw, width=640, height=360, format=BGR" ! fakesink -v

One input multiple output

The following example configures vvas_xabrscaler for one input and three outputs. The input is 1920 x 1080 resolution in NV12 format. There are three output formats:

• 1280 x 720 in BGR format

• 300 x 300 in RGB format

• 640 x 480 in NV12 format

gst-launch-1.0 videotestsrc num-buffers=100 \
! "video/x-raw, width=1920, height=1080, format=NV12, framerate=60/1" \
! vvas_xabrscaler xclbin-location="/usr/lib/dpu.xclbin" kernel-name=v_multi_scaler:v_multi_scaler_1 name=sc sc.src_0 \
! queue \
! "video/x-raw, width=1280, height=720, format=BGR" \
! fakesink sc.src_1 \
! queue \
! "video/x-raw, width=300, height=300, format=RGB" \
! fakesink sc.src_2 \
! queue \
! "video/x-raw, width=640, height=480, format=NV12" \
! fakesink -v

vvas_xinfer

GStreamer inference vvas_xinfer plugin performs inferencing on video frames with the help of Vitis AI library and prepares tree like metadata in GstInferenceMeta object and attaches the same to input GstBuffer. This plugin triggers optional preprocessing (scale/crop & etc.) operations with the help of VVAS scaler kernel library (which is on top of Xilinx’s multiscaler IP) on incoming video frames before calling VVAS inference kernel library. vvas_xinfer plugin’s input capabilities are influenced by preprocessing library input capabilities and Vitis AI library capabilities (Vitis-AI does software scaling). This plugin disables preprocessing whenever the input resolution & color format is same as inference engine’s resolution & color format. If preprocessing is enabled and vvas_xinfer plugin is receiving non-CMA memory frames, then data copy will be made to ensure CMA frames goes to preprocessing engine. The main advantage of this plugin is users/customers can realize inference cascading use cases with ease.

Input and Output

• Accepts buffers with GRAY8/ NV12/ BGR/ RGB/ YUY2/ r210/ v308/ GRAY10_LE32/ ABGR/ ARGB color formats on input GstPad & output GstPad.

• Attaches GstInferenceMeta metadata to output GstBuffer

Control Parameters

Control Parameters

Property Name	Type	Range	Default	Description
dynamic-config	String	N/A	Null	String contains dynamic json configuration of inference accelration library
infer-config	String	N/A	Null	location of the inference kernel library configuration file in json format
preprocess-config	String	N/A	Null	location of the kernels config file in json format

infer-config json members

Json key	Item	Item description
vvas-library-repo	Description	VVAS libraries repository path to look for kernel libraries by VVAS GStreamer plugin
	Value type	String
	Mandatory/Optional	Optional
	Default value	/usr/lib in Embedded platforms /opt/xilinx/vvas/lib in PCIe platforms
inference-level	Description	Inference level in cascaded inference use case. e.g. Object detection ML (level-1) followed by object classification (level-2) on detected objects
	Value type	Integer
	Mandatory/Optional	Optional
	Default value	1
low-latency	Description	Parameter to enable/disable low-latency mode in vvas_xinfer and it is useful only when inference-level > 1. If enabled, then vvas_xinfer plugin will not wait for batch-size frames to be accumulated to reduce latency. If disabled, inference engine can work at maximum throughput.
	Value type	Boolean
	Mandatory/Optional	Optional
	Default value	false
inference-max-queue	Description	Maximum number of input frames those can be queued inside the plugin. When low-latency is disabled, vvas_xinfer plugin will wait for inference-max-queue buffers until batch-size is accumulated
	Value type	Integer
	Mandatory/Optional	Optional
	Default value	batch-size
attach-ppe-outbuf	Description	Attaches output of preprocessing library to GstInferenceMeta to avoid redoing of the preprocessing if required.
	Value type	Boolean
	Mandatory/Optional	Optional
	Default value	False
kernel	Description	Kernel object provides information about an VVAS kernel library configuration and kernel library name
	Value type	JSON Object
	Mandatory/Optional	Mandatory
	Default value	None
	Object Members	members of kernel JSON object are mentioned below

infer-config::kernel json members

JSON key	Item	Description
library-name	Description	Name of inference kernel library to be loaded for inferencing
	Value type	String
	Mandatory/Optional	Mandatory
	Default value	NULL
config	Description	Inference kernel specific configuration
	Value type	JSON object
	Mandatory/Optional	Mandatory
	Default value	None
	Object members	Contains members specific to inference kernel library. See vvas_xdpuinfer acceleration library for more information

preprocess-config json members

Table 5 preprocess-config json members

Json key	Item	Item description
xclbin-location	Description	Location of xclbin which contains scaler IP to program FPGA device based on device-index property
	Value type	String
	Mandatory/Optional	Mandatory
	Default value	NULL
vvas-library-repo	Description	VVAS libraries repository path to look for kernel libraries by VVAS GStreamer plugin
	Value type	String
	Mandatory/Optional	Optional
	Default value	/usr/lib in Embedded platforms /opt/xilinx/vvas/lib in PCIe platforms
device-index	Description	Device index on which scaler IP is present
	Value type	Integer
	Mandatory/Optional	Mandatory in PCIe platforms In embedded platforms, device-index is not an applicable option as it is always zero
	Default value	-1 in PCIe platforms 0 in Embedded platforms
kernel	Description	Kernel object provides information about an VVAS kernel library configuration and kernel library name
	Value type	JSON Object
	Mandatory/Optional	Mandatory
	Default value	None
	Object Members	members of kernel JSON object are mentioned below

preprocess-config::kernel json members

Table 6: preprocess-config::kernel json members

JSON key	Item	Description
library-name	Description	Name of kernel library to be loaded for inferencing
	Value type	String
	Mandatory/Optional	Mandatory
	Default value	NULL
kernel-name	Description	Name of the preprocessing kernel. Syntax : “<kernel_name>:<instance_name>”
	Value type	String
	Mandatory/Optional	Mandatory
	Default value	NULL
config	Description	preprocess kernel specific configuration
	Value type	JSON object
	Mandatory/Optional	Mandatory
	Default value	None
	Object members	Contains members specific to preprocess kernel library

• Example infer-json file:

{
   "vvas-library-repo": "/usr/lib/",
   "inference-level":1,
   "inference-max-queue":30,
   "attach-ppe-outbuf": false,
   "low-latency":false,
   "kernel" : {
       "library-name":"libvvas_xdpuinfer.so",
       "config": {
           "model-name" : "yolov3_voc_tf",
           "model-class" : "YOLOV3",
           "model-format" : "RGB",
           "batch-size":1,
           "model-path" : "/usr/share/vitis_ai_library/models/",
           "run_time_model" : false,
           "need_preprocess" : false,
           "performance_test" : false,
           "debug_level" : 0
       }
   }
}

• Example preprocess-config json file:

{
   "xclbin-location":"/usr/lib/dpu.xclbin",
   "vvas-library-repo": "/usr/lib",
   "device-index": 0,
   "kernels" :[
       {
           "kernel-name": "v_multi_scaler:v_multi_scaler_1",
           "library-name": "libvvas_xpreprocessor.so",
           "config": {
              "alpha_r" : 0.0,
              "alpha_g" : 0.0,
              "alpha_b" : 0.0,
              "beta_r" : 0.25,
              "beta_g" : 0.25,
              "beta_b" : 0.25,
              "cascade" : 1,
              "debug_level" : 0
           }
       }
   ]

Note

When user wants to perform hardware accelerated pre-processing on input frame using vvas_xinfer plug-in, then the pre-processing parameters provided in preprocess-config json file must be checked for correctness. The pre-processing parameters for the models are provided in the prototxt file for the models. For few models these parameters needs to be modified from what it is in the prototxt file. User must provide these modified values in the preprocess-config json file. There are few steps mentioned below to know which models need changes.

To determine the pre-processing parameters for a model, follow the steps mentioned below:

• Get the algorithmic scale vector, i.e. “scale” fields for each channel R,G,B mentioned in the model prototxt file. If the model expects input image in RGB format, then the first field will corresponds to Channel R, next will corresponds to Channel G, and the last one will be for Channel B. Example prototxt file contents are mentioned below:

model {
        name: "yolov3_voc_416"
        kernel {
            name: "yolov3_voc_416"
            mean: 0.0
            mean: 0.0
            mean: 0.0
            scale: 0.00390625
            scale: 0.00390625
            scale: 0.00390625
}

• Calculate the inner scale value from the “fixpos” field of the “input_tensor” by executing the command at command prompt on the target board in case of Embedded platform.

xdputil xmodel -l path-to-xmodel.xmodel

you can get something like:

{
 "subgraphs":[
     {
         "name":"subgraph_data",
         "device":"USER"
     },
     {
        "name":"subgraph_conv1_1",
         "device":"DPU",
         "fingerprint":"0x1000020f6014407",
         "DPU Arch":"DPUCZDX8G_ISA0_B4096_MAX_BG2",
         "input_tensor":[
             {
                 "name":"data_fixed",
                 "shape":"[1, 360, 480, 3]",
                 "fixpos":-1
             }
         ],


 inner scale = 2^fixpos
             = 2^-1
             = 0.5

• Multiply the “scale” vector values from prototxt file with the inner scale calculated above. Provide these values in the preprocess-config json file. User need to be careful while assigning the value to the correct channel field.

• Example Simple inference (YOLOv3) pipeline which takes NV12 YUV file (test.nv12) as input is described below:

gst-launch-1.0 filesrc location=<test.nv12> ! videoparse width=1920 height=1080 format=nv12 ! \
vvas_xinfer preprocess-config=yolo_preproc.json infer-config=yolov3_voc.json ! fakesink -v

• Example cascade inference (YOLOv3+Resnet18) pipeline which takes NV12 YUV file (test.nv12) as input is described below:

gst-launch-1.0 filesrc location=<test.nv12> ! videoparse width=1920 height=1080 format=nv12 ! \
vvas_xinfer preprocess-config=yolo_preproc.json infer-config=yolov3_voc.json ! queue ! \
vvas_xinfer preprocess-config=resnet_preproc.json infer-config=resnet18.json ! fakesink -v

Infrastructure Plug-ins and Acceleration Software Libraries

Infrastructure plug-ins are generic plug-ins that interact with the acceleration kernel through a set of APIs exposed by an acceleration software library corresponding to that kernel. Infrastructure plug-ins abstract the core/common functionality of the GStreamer framework (for example: caps negotiation and buffer management).

Table 5: Infrastructure Plug-ins

Infrastructure Plug-ins	Function
vvas_xfilter	Plug-in has one input, one output Support Transform, passthrough and inplace transform operations
vvas_xmultisrc	Plug-in support one input and multiple output pads. Support transform operation

Note

Though one input and one output kernel can be integrated using any of the two infrastructure plug-ins, we recommend using vvas_xfilter plugin for one input and one output kernels.

Acceleration software libraries control the acceleration kernel, like register programming, or any other core logic required to implement the functions. Acceleration software libraries expose a simplified interface that is called by the GStreamer infrastructure plug-ins to interact with the acceleration kernel. The following table lists the acceleration software libraries developed to implement specific functionality. These libraries are used with one of the infrastructure plug-ins to use the functionality a GStreamer-based application. Example pipelines with GStreamer infrastructure plug-ins and acceleration software libraries are covered later in this section.

Table 6: Acceleration Software Libraries

Acceleration Software Library	Function
vvas_xdpuinfer	Library based on Vitis AI to control DPU kernels for machine learning.
vvas_xboundingbox	Library to draw a bounding box and labels on the frame using OpenCV.

The GStreamer infrastructure plug-ins are available in the vvas-gst-plugins repository/ folder. The following section describes each infrastructure plug-in.

vvas_xfilter

The GStreamer vvas_xfilter is an infrastructure plug-in that is derived from GstBaseTransform. It supports one input pad and one output pad. The vvas_xfilter efficiently works with hard-kernel/soft-kernel/software (user-space) acceleration software library types as shown in the following figure.

This plug-in can operate in the following three modes.

• Passthrough: Useful when the acceleration software library does not need to alter the input buffer.

• In-place: Useful when the acceleration software library needs to alter the input buffer.

• Transform: In this mode, for each input buffer, a new output buffer is produced.

You must set the mode using the JSON file. Refer to JSON File Schema for information related to the kernels-config property.

The vvas_xfilter plug-in takes configuration file as one of the input properties, kernels- config. This configuration file is in JSON format and contains information required by the kernel. During initialization, the vvas_xfilter parses the JSON file and performs the following tasks:

• Finds the VVAS acceleration software library in the path and loads the shared library.

• Understands the acceleration software library type and prepares the acceleration software library handle (VVASKernel) to be passed to the core APIs.

Input and Output

The vvas_xfilter accepts the buffers with the following color formats on input GstPad and output GstPad.

• GRAY8

• NV12

• BGR

• RGB

• YUY2

• r210

• v308

• GRAY10_LE32

• ABGR

• ARGB

The formats listed are the Xilinx IP supported color formats. To add other color formats, update the vvas_kernel.h and vvas_xfilter plug-ins.

Control Parameters and Plug-in Properties

The following table lists the GObject properties exposed by the vvas_xfilter. Most of them are only available in PCIe supported platforms.

Table 6: GObject Properties

Property Name	Platforms Supported	Type	Range	Default	Description
dynamic-config	PCIe and Embedded	String	N/A	Null	JSON formatted string contains kernel specific configuration for run time changes
dev-idx	PCIe only	Integer	0 to 31	0	Device used for kernel processing, xclbin download.
kernels-config	PCIe and Embedded	String	N/A	NULL	JSON configuration file path based on VVAS acceleration software library requirements. Refer to the JSON File Schema
sk-cur-idx	PCIe only	Integer	0 to 31	0	Softkernel current index to be used for executing job on device.
reservation-id	PCIe only	Integer	0 to 1024	0	Reservation ID provided by the Xilinx resource manager (XRM).

JSON Format for vvas_xfilter Plug-in

The following table provides the JSON keys accepted by the GStreamer vvas_xfilter plug-in.

Table 7: Root JSON Object Members

JSON Key	Item	Item Description
xclbin-location	Description	The location of the xclbin that is used to program the FPGA device.
	Value type	String
	Mandatory or optional	Conditionally mandatory: If the VVAS acceleration software library is developed for hard-kernel IP and soft-kernel, then the xclbin-location is mandatory. If the VVAS acceleration software library is developed for a software library (e.g., OpenCV), then the xclbin-location is not required
	Default value	NULL
vvas-library-repo	Description	This is the VVAS libraries repository path to look for acceleration software libraries by the VVAS GStreamer plug-in.
	Value type	String
	Mandatory or optional	Optional
	Default value	/usr/lib
element-mode	Description	GStreamer element mode to operate. Based on your requirement, choose one of the following modes: Passthrough: In this mode, element does not want to alter the input buffer. Inplace: In this mode, element wants to alter the input buffer itself instead of producing new output buffers. Transform: In this mode, element produces a output buffer for each input buffer.
	Value type	Enum
	Mandatory or optional	Mandatory
kernels	Description	This is the array of kernel objects. Each kernel object provides information about an VVAS acceleration software library configuration. The vvas_xfilter only takes the first kernel object in the kernel array.
	Value type	Array of objects
	Mandatory or optional	Mandatory
	Default value	None
	Minimum value	1
	Maximum value	10
	Object members	Refer to Kernel JSON Object

The information in the following table is specific to the kernel chosen.

Table 8: Kernel JSON Object Members

JSON Key	Item	Item Description
library-name	Description	The name of the VVAS acceleration software library to be loaded by the VVAS GStreamer plug-ins. The absolute path of the kernel library is formed by the pre-pending vvas-library-repo path.
	Value type	String
	Mandatory or optional	Mandatory
	Default value	None
kernel-name	Description	The name of the IP/kernel in the form of <kernel name>:<instance name>
	Value type	String
	Mandatory or optional	Optional
	Default value	None
config	Description	Holds the configuration specific to the VVAS acceleration software library. VVAS GStreamer plug-ins do not parse this JSON object, instead it is sent to the acceleration software library.
	Value type	Object
	Mandatory or optional	Optional
	Default value	None
softkernel	Description	Contains the information specific to the soft-kernel. This JSON object is valid only for the PCIe based platforms.
	Value type	Object
	Mandatory or optional	Mandatory if kernel library is written for soft-kernel.
	Default value	None
	Members	Not required for embedded platforms.

For vvas_xfilter implementation details, refer to vvas_xfilter source code

Example JSON Configuration Files

JSON File for Vitis AI API-based VVAS Acceleration Software Library

The following JSON file is for pure software-based acceleration, it does not involve any kernel for acceleration. However, the Vitis AI API based DPU is a special case, where the DPU hardware kernel is controlled by Vitis AI. The acceleration software library calls the Vitis AI API calls and it is implemented as a pure software acceleration software library. There is no need to provide the path of the xclbin in the JSON file.

{
   "vvas-library-repo": "/usr/lib/",
   "element-mode":"inplace",
   "kernels" :[
      {
         "library-name":"libvvas_xdpuinfer.so",
         "config": {
            "model-name" : "densebox_320_320",
            "model-class" : "FACEDETECT",
            "model-format": ""BGR",
            "model-path" : "/usr/share/vitis_ai_library/models/",
            "run_time_model" : false,
            "need_preprocess" : true,
            "performance_test" : true,
            "max_num" : -1,
            "prob_cutoff" : 0.0,
            "debug_level" : 1
         }
      }
   ]
}

JSON File for a Hard Kernel

The following JSON file uses vvas_xfilter to control multi-scaler IP (hard-kernel). The acceleration software library accessing the register map is libvvas_xcrop.so.

{
   "xclbin-location":"/usr/lib/dpu.xclbin",
   "vvas-library-repo": "/usr/lib/",
   "element-mode":"passthrough",
   "kernels" :[
       {
          "kernel-name":"v_multi_scaler:v_multi_scaler_1",
          "library-name":"libvvas_xcrop.so",
          "config": {
          }
       }
   ]
}

Example Pipelines

This section covers the GStreamer example pipelines using the vvas_xfilter infrastructure plug-in and several acceleration software libraries. This section covers the bounding box functionality and the machine learning functions using the vvas_xfilter plug-in.

• The bounding box functionality is implemented in the vvas_xboundingbox acceleration software library that is controlled by the vvas_xfilter plug-in.

• Machine learning using the DPU is implemented by the vvas_xdpuinfer acceleration software library that is called by the vvas_xfilter plug-in.

Bounding Box Example

This section describes how to draw a bounding box and label information using the VVAS infrastructure plug-in vvas_xfilter and the vvas_xboundingbox acceleration software library. The vvas_xboundingbox interprets machine learning inference results from the vvas_xdpuinfer acceleration software library and uses an OpenCV library to draw the bounding box and label on the identified objects.

For vvas_xboundingbox implementation details, refer to vvas_xboundingbox source code

Prerequisites

There are a few requirements before start running bounding box examples. Make sure these prerequisites are met.

• Installation of OpenCV: The vvas_xboundingbox uses OpenCV for the graphics back-end library to draw the boxes and labels. Install the libopencv-core-dev package (the preferred version is 3.2.0 or later).

JSON File for vvas_xboundingbox

This section describes the JSON file format and configuration parameters for the bounding box acceleration software library. The GStreamer vvas_xfilter plug-in used in the inplace mode. Bounding box and labels are drawn on identified objects on the incoming frame. Bounding box functionality is implemented in the libvvas_xboundingbox.so acceleration software library.

The following example is of a JSON file to pass to the vvas_xfilter.

{
   "xclbin-location":"/usr/lib/dpu.xclbin",
   "vvas-library-repo": "/usr/local/lib/vvas",
   "element-mode":"inplace",
   "kernels" :[
      {
         "library-name":"libvvas_xboundingbox.so",
         "config": {
            "font_size" : 0.5,
            "font" : 3,
            "thickness" : 2,
            "debug_level" : 2,
            "label_color" : { "blue" : 0, "green" : 0, "red" : 0 },
            "label_filter" : [ "class", "probability" ],
            "classes" : [
               {
                  "name" : "car",
                  "blue" : 255,
                  "green" : 0,
                  "red" : 0
               },
               {
                  "name" : "person",
                  "blue" : 0,
                  "green" : 255,
                  "red" : 0
               },
               {
                  "name" : "bicycle",
                  "blue" : 0,
                  "green" : 0,
                  "red" : 255
               }
            ]
         }
      }
   ]
}

Various configuration parameters of the bounding box acceleration software library are described in the following table.

Table 9: vvas_xboundingbox Parameters

Parameter	Expected Values	Description
debug_level	0: LOG_LEVEL_ERROR 1: LOG_LEVEL_WARNING 2: LOG_LEVEL_INFO 3: LOG_LEVEL_DEBUG	Enables the log levels. There are four log levels listed in the expected values column.
font	0 to 7	Font for the label text. 0: Hershey Simplex 1: Hershey Plain 2: Hershey Duplex 3: Hershey Complex 4: Hershey Triplex 5: Hershey Complex Small 6: Hershey Script Simplex 7: Hershey Script Complex
font_size	0.5 to 1	Font fraction scale factor that is multiplied by the font-specific base size.
thickness	Integer 1 to 3	The thickness of the line that makes up the rectangle. Negative values, like -1, signify that the function draws a filled rectangle. The recommended value is between 1 and 3.
label_color	{ “blue” : 0, “green” : 0, “red” : 0 }	The color of the label is specified.
label_filter	[ “class”, “probability” ]	This field indicates that all information printed is the label string. Using “class” alone adds the ML classification name. For example, car, person, etc. The addition of “probability” in the array adds the probability of a positive object identification.
classes	{ “name” : “car”, “blue” : 255, “green” : 0, “red” : 0 }	This is a filtering option when using the vvas_xboundingbox. The bounding box is only drawn for the classes that are listed in this configuration. Other classes are ignored. For instance, if “car”, “person”, “bicycle” is entered under “classes”, then the bounding box is only drawn for these three classes, and other classes like horse, motorbike, etc. are ignored. The expected value columns show an example of how each class should be described. All objects in this example, by class, are drawn using the color combination listed. The class names in this list matches the class names assigned by the vvas_xdpuinfer. Otherwise, the bounding box is not drawn. For face detect, keep the “classes” array empty.
display_level	Integer 0 to N 0 => all levels N => specific level	display bounding box of one particular level or all levels

An example of using a bounding box along with the machine learning plug-in is shown in the Multi Channel ML Tutorial.

Machine Learning Example

This section discusses how machine learning solutions can be implemented using the VVAS infrastructure vvas_xfilter plug-in and the vvas_xdpuinfer acceleration software library.

The vvas_xdpuinfer is the acceleration software library that controls the DPU through the Vitis AI interface. The vvas_xdpuinfer does not modify the contents of the input buffer. The input buffer is passed to the Vitis AI model library that generates the inference data. This inference data is then mapped into the VVAS meta data structure and attached to the input buffer. The same input buffer is then pushed to the downstream plug-in.

For vvas_xdpuinfer implementation details, refer to vvas_xdpuinfer source code

Prerequisites

There are a few requirements to be met before you can start running the machine learning examples.

Model Information

The model directory name must match with the ELF and prototxt files inside the model directory. The model directory should contain:

• model-name.elf/model-name.xmodel and model-name.prototxt file.

• label.json file containing the label information, if the models generate labels.

The following is an example of the model directory (yolov2_voc), which contains the yolov2_voc.xmodel and yolov2_voc.prototxt files along with the label.json file.

xclbin Location

By default, the Vitis AI interface expects xclbin to be located at /usr/lib/ and the xclbin is called dpu.xclbin. Another option is to use the environment variable XLNX_VART_FIRMWARE to change the path and the corresponding path can be updated in the JSON file. That is, export XLNX_VART_FIRMWARE=/where/your/dpu.xclbin.

Input Image

The vvas_xdpuinfer works with raw BGR and RGB images as required by the model. Make sure you have specified correct color format in model-format field in json file. The exact resolution of the image to vvas_xdpuinfer must be provided, it is expected by the model. There is a performance loss if a different resolution of the BGR image is provided to the vvas_xdpuinfer, because resizing is done on the CPU inside the Vitis AI library.

JSON File for vvas_xdpuinfer

The following table shows the JSON file format and configuration parameters for vvas_xdpuinfer.

Table 10: JSON File for vvas_xdpuinfer

Parameter	Type	Expected Values	Default	Description
xclbin-location	string	This field is not needed for dpuinfer	NULL	By default, Vitis AI expects xclbin to be located at /usr/lib/ and xclbin is called “dpu.xclbin”. The environment variable XLNX_VART_FIRMWARE could also be used to change the path. For example, export XLNX_VART_FIRMWARE=/where/your/dpu.xclbin. Alternatively, you may also provide the xclbin path in /etc/vart.conf. In this release, /etc/vart.conf has “/media/sd-mmcblk0p1/dpu.xclbin”
vvas-library-repo	string	/usr/local/lib/vvas/	usr/lib/	This is the path where the vvas_xfilter will search the acceleration software library. The kernel name is specified in the library-name parameter of the JSON file.
element-mode	string	inplace	None	Because the input buffer is not modified by the ML operation, but the metadata generated out of an inference operation needs to be added/appended to the input buffer, the GstBuffer is writable. The vvas_xfilter is configured in inplace mode
kernels	N/A	N/A	N/A	The JSON tag for starting the kernel specific configurations.
kernel-name	string	N/A	NULL	The name and instance of a kernel separated by “:”
library-name	string	N/A	NULL	Acceleration software library name for the kernel. It is appended to the vvas-l ibrary-repo for an absolute path.
config	N/A	N/A	N/A	The JSON tag for kernel-specific configurations depending on the acceleration software library.
model-name	string	resnet50	N/A	Name string of the machine learning model to be executed. The name string should be same as the name of the directory available in model -path parameter file. If the name of the model ELF file is resnet50.elf, then the model-name is resnet50 in the JSON file. The ELF file present in the specified path model-path of the JSON file.
model-class	string	YOLOV3 FACEDETECT CLASSIFICATION SSD REFINEDET TFSSD YOLOV2	N/A	Class of some model corresponding to model. Some examples are shown below: YOLOV3: yolov3_adas_pruned_0_9, yolov3_voc, yolov3_voc_tf FACEDETECT: densebox_320_320, densebox_640_360 CLASSIFICATION: resnet18, resnet50, resnet_v1_50_tf
model-format	string	RGB/BGR	N/A	Image color format required by model.
model-path	string	/usr/share/vitis_ai_library/models/	N/A	Path of the folder where the model to be executed is stored.
run_time_model	Boolean	True/False	False	If there is a requirement to change the ML model at the frame level, then set this flag to true. If this parameter is set to true then vvas_xdpuinfer will read the model name and class from the incoming input metadata and execute the same model found in the path specified in the model-path. The model-name and model-class parameter of the JSON file are not required when enabling this parameter.
need_preprocess	Boolean	True/False	True	If need_preprocess = true: Normalize with mean/scale through the Vitis AI Library If need_preprocess = false: Normalize with mean/scale is performed before feeding the frame to vvas_xdpuinfer. The Vitis AI library does not perform these operations.
performance_test	Boolean	True/False	False	Enable performance test and corresponding flops per second (f/s) display logs. Calculates and displays the f/s of the standalone DPU after every second.
debug_level	integer	0 to 3	1	Used to enable log levels. There are four log levels for a message sent by the kernel library code, starting from level 0 and decreasing in severity till level 3 the lowest log-level identifier. When a log level is set, it acts as a filter, where only messages with a log-level lower than it, (therefore messages with an higher severity) are displayed. 0: This is the highest level in order of severity: it is used for messages about critical errors, both hardware and software related. 1: This level is used in situations where you attention is immediately required. 2: This is the log level used for information messages about the action performed by the kernel and output of model. 3: This level is used for debugging.

Model Parameters

The Vitis AI library provides a way to read model parameters by reading the configuration file. It facilitates uniform configuration management of model parameters. The configuration file is located in the model-path of the JSON file with [model_name].prototxt. These parameters are model specific. For more information on model parameters, refer to the Vitis AI Library User Guide (UG1354).

Example GStreamer Pipelines

This section describes a few example GStreamer pipelines.

Classification Example Using Resnet50

The following pipeline performs ML using a Resnet50 model. DPU configuration uses kernels- config = ./json_files/kernel_resnet50.json for the vvas_xdpuinfer. The output of the vvas_xfilter is passed to fakesink along with the metadata.

The GStreamer command for the example pipeline:

gst-launch-1.0 filesrc location="<PATH>/001.bgr" blocksize=150528 numbuffers=1
! videoparse width=224 height=224 framerate=30/1 format=16
! vvas_xfilter name="kernel1" kernels-config="<PATH>/kernel_resnet50.json"
! fakesink

The JSON file for the vvas_xdpuinfer to execute resnet50 model based classification pipeline is described below.

{
   "vvas-library-repo": "/usr/local/lib/vvas/",
   "element-mode":"inplace",
   "kernels" :[
      {
         "library-name":"libvvas_xdpuinfer.so",
         "config": {
            "model-name" : "resnet50",
            "model-class" : "CLASSIFICATION",
            "model-format : "BGR"
            "model-path" : "/usr/share/vitis_ai_library/models/",
            "run_time_model" : false,
            "need_preprocess" : true,
            "performance_test" : true,
            "debug_level" : 2
         }
      }
   ]
}

Note

If “need_preprocess” = false, then pre-processing operations like, Normalization, scaling must be performed on the frame before feeding to vvas_xfilter/vvas_xdpuinfer otherwise results may not be as expected.

vvas_xmultisrc

The GStreamer vvas_xmultisrc plug-in can have one input pad and multiple-output pads. This single plug-in supports multiple acceleration kernels, each controlled by a separate acceleration software library.

For vvas_xmultisrc implementation details, refer to vvas_xmultisrc source code

Input and Output

Input and output accept buffers with the following color formats on input GstPad and output GstPad.

• GRAY8

• NV12

• BGR

• RGB

• YUY2

• r210

• v308

• GRAY10_LE32

• ABGR

• ARGB

The formats listed are the Xilinx IP supported color formats. To add other color formats, update the vvas_kernel.h and vvas_xfilter plug-ins.

Control Parameters and Plug-in Properties

Table 11: Plug-in Properties

Property Name	Type	Range	Default	Description
kconfig	String	N/A	NULL	Path of the JSON configuration file based on the VVAS acceleration software library requirements. For further information, refer to JSON File Schema
dynamic-config	String	N/A	NULL

JSON File for vvas_xmultisrc

This section covers the format of JSON file/string to be passed to the vvas_xmultisrc plug-in.

Example JSON File

The following example file describes how to specify two kernels that are being controlled by the single instance of the vvas_xmultisrc plug-in. Modify this json file as per your kernels and acceleration software library. The next section describes each field in this example file.

{
   "xclbin-location":"/usr/lib/binary_1.xclbin",
   "vvas-library-repo": "/usr/lib/",
   "kernels" :[
      {
         "kernel-name":"resize:resize_1", <------------------ kernel 1
         "library-name":"libvvas_xresize.so",
         "config": {
            x : 4,
            y : 7
         }
      }
      {
         "kernel-name":"cvt_rgb:cvt_rgb_1", <-------------- kernel 2
         "library-name":"libcvt_bgr.so",
         "config": {
            name = "xilinx",
            value = 98.34
         }
      }
   ]
}

Table 12: JSON Properties

Property Name	Type	Range	Default	Description
xclbin-location	String	N/A	NULL	The path of xclbin including the xclbin name. The plug-in downloads this xclbin and creates an XRT handle for memory allocation and programming kernels.
vvas-library-repo	String	N/A	/usr/lib	The library path for the VVAS repository for all the acceleration software libraries.
kernels	N/A	N/A	N/A	The JSON tag for starting the kernel specific configurations.
kernel-name	String	N/A	NULL	Name and instance of a kernel separated by “:” as mentioned in xclbin.
library-name	String	N/A	NULL	The acceleration software library name for the kernel. This is appended to vvas-library-repo for an absolute path.
config	N/A	N/A	N/A	The JSON tag for kernel specific configurations that depends on the acceleration software library.

Source Pad Naming Syntax

For single output pad naming is optional. For multiple pads, the source pads names shall be as mentioned below, assuming the name of the plug-in as sc.

sc.src_0, sc.src_1 …..sc.src_n

Example Pipelines

Single Output Pad

The following example demonstrates the vvas_xmultisrc plug-in configured for one input and one output. A test video pattern is generated by the videotestrc plug-in and passed to the vvas_xmultisrc plug-in. Depending on the kernel being used, vvas_xmultisrc uses kernel.json to configure the kernel for processing the input frame before passing it to the fakesink.

gst-launch-1.0 videotestsrc
! "video/x-raw, width=1280, height=720, format=BGR"
! vvas_xmultisrc kconfig="/root/jsons/<kernel.json>"
! "video/x-raw, width=640, height=360, format=BGR"
! fakesink -v

The following is an example kernel.json file having mean_value and use_mean as kernel configuration parameters. Modify this as per your kernel requirements.

{
   "xclbin-location": "/usr/lib/dpu.xclbin",
   "vvas-library-repo": "/usr/lib/vvas",
   "kernels": [
      {
         "kernel-name": "<kernel-name>",
         "library-name": "libvvas_xresize_bgr.so",
         "config": {
         "use_mean": 1,
         "mean_value": 128
         }
      }
   ]
}

GstVvasBufferPool

The GStreamer VVAS buffer pool holds the pool of video frames allocated using the GStreamer allocator object. It is derived from the GStreamer base video buffer pool object (GstVideoBufferPool).

The VVAS buffer pool:

• Allocates buffers with stride and height alignment requirements. (e.g., the video codec unit (VCU) requires the stride to be aligned with 256 bytes and the height aligned with 64 bytes)

• Provides a callback to the GStreamer plug-in when the buffer comes back to the pool after it is used.

The following API is used to create the buffer pool.

GstBufferPool *gst_vvas_buffer_pool_new (guint stride_align, guint
elevation_align)

Parameters:
   stride_align - stride alignment required
   elevation_align - height alignment required

Return:
   GstBufferPool object handle

Plug-ins can use the following API to set the callback function on the VVAS buffer pool and the callback function is called when the buffer arrives back to the pool after it is used.

Buffer release callback function pointer declaration:
typedef void (*ReleaseBufferCallback)(GstBuffer *buf, gpointer user_data);

void gst_vvas_buffer_pool_set_release_buffer_cb (GstVvasBufferPool *xpool,
ReleaseBufferCallback release_buf_cb, gpointer user_data)

Parameters:
   xpool - VVAS buffer pool created using gst_vvas_buffer_pool_new
   release_buf_cb - function pointer of callback
   user_data - user provided handle to be sent as function argument while
calling release_buf_cb()

Return:
    None

GstVvasAllocator

The GStreamer VVAS allocator object is derived from an open source GstAllocator object designed for allocating memory using XRT APIs. The VVAS allocator is the backbone to the VVAS framework achieving zero-copy (wherever possible).

Allocator APIs

GStreamer plug-in developers can use the following APIs to interact with the VVAS allocator. To allocate memory using XRT, the GStreamer plug-ins and buffer pools require the GstAllocator object provided by the following API.

GstAllocator* gst_vvas_allocator_new (guint dev_idx, gboolean need_dma)

Parameters:
   dev_idx - FPGA Device index on which memory is going to allocated
   need_dma - will decide memory allocated is dmabuf or not

Return:
   GstAllocator handle on success or NULL on failure

..note:: In PCIe platforms, the DMA buffer allocation support is not available. This means that the need_dma argument to gst_vvas_allocator_new() API must be false.

Use the following API to check if a particular GstMemory object is allocated using GstVvasAlloctor.

gboolean gst_is_vvas_memory (GstMemory *mem)

Parameters:
   mem - memory to be validated

Return:
   true if memory is allocated using VVAS Allocator or false

When GStreamer plug-ins are interacting with hard-kernel IP or soft-kernel, the plug-ins need physical memory addresses on an FPGA using the following API.

guint64 gst_vvas_allocator_get_paddr (GstMemory *mem)

Parameters:
   mem - memory to get physical address

Return:
   valid physical address on FPGA device or 0 on failure

Use the following API when plug-ins need an XRT buffer object (BO) corresponding to an VVAS memory object.

guint gst_vvas_allocator_get_bo (GstMemory *mem)

Parameters:
   mem - memory to get XRT BO

Return:
   valid XRT BO on success or 0 on failure

JSON Schema

This section covers the JSON schema for the configuration files used by the infrastructure plug-ins. For more details, refer to JSON Schema

VVAS Inference Meta Data

This section covers details about inference meta data structure used to store meta data. For more details, refer to VVAS Inference Meta Data