Skip to content
This repository has been archived by the owner on Aug 20, 2024. It is now read-only.

Latest commit

 

History

History
486 lines (409 loc) · 22.2 KB

flex_170_gpu.md

File metadata and controls

486 lines (409 loc) · 22.2 KB
Raw

Intel® Data Center GPU Flex Series 170 Guide

| Installation | Object Detection | Object Classification |

Note: These instructions assume that your host environment and firmware have already been installed.

Actual results will vary based on configuration and the examples below are for illustration purposes only.

Note: For best results enable performance mode for the CPU.

echo performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

Installation

Use the following commands to enable discrete graphics support.

  1. Clone Repository

    git clone https://github.com/dlstreamer/pipeline-zoo.git pipeline-zoo

  2. Build Pipeline Zoo Environment

    ./pipeline-zoo/tools/docker/build.sh --platform dgpu

    Expected Output

    Removing intermediate container d186e3cdb2e6
---> 84423666f6d7
Successfully built 84423666f6d7
Successfully tagged dlstreamer-pipeline-zoo-dgpu:latest

    Note: If you need to use a different base image (for example with updated drivers or local modifications) pass the image via the --base argument.

    ./pipeline-zoo/tools/docker/build.sh --platform dgpu --base <IMAGE_NAME_HERE>

  3. Launch Pipeline Zoo

    ./pipeline-zoo/tools/docker/run.sh --platform dgpu

    Expected Output

    Running Media Analytics Pipeline Zoo Image: 'dlstreamer-pipeline-zoo-dgpu'
   Environment: '-e no_proxy -e https_proxy -e http_proxy -e DISPLAY  -e DEVICE=/dev/dri/renderD128'
   Volume Mounts: '-v /home/gta/.netrc:/root/.netrc -v /home/pipeline-zoo:/home/pipeline-zoo/ -v /tmp:/tmp -v /var/tmp:/var/tmp -v /dev:/dev -v /lib/modules:/lib/modules -v /home/gta/.Xauthority:/root/.Xauthority '
   Name: 'dlstreamer-pipeline-zoo-dgpu'
...

Object Detection Pipelines

stateDiagram
    direction LR 
    state Object-Detection {
    direction LR
    state video_source {
	direction LR
		demux --> parse 
    }
   
    state detect {
	direction LR
		scale --> csc
		csc --> inference
		inference --> tensors_to_objects
    }
    
    media --> video_source
    video_source --> decode
    decode --> detect
    detect --> objects
}
Loading 
+----------------------------------------------+-----------------------+----------------------------+------------+
| Pipeline                                     | Task                  | Models                     | Runners    |
+==============================================+=======================+============================+============+
| od-h264-yolov5s-416                          | object-detection      | yolov5s-416                | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h264-ssd-mobilenet-v1-coco                | object-detection      | ssd_mobilenet_v1_coco      | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h264-yolo-v4-tf                           | object-detection      | yolo-v4-tf                 | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h265-ssd-mobilenet-v1-coco                | object-detection      | ssd_mobilenet_v1_coco_INT8 | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h265-ssdlite-mobilenet-v2                 | object-detection      | ssdlite_mobilenet_v2       | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h265-yolo-v4-tf                           | object-detection      | yolo-v4-tf                 | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h264-yolo-v4-tf-416                       | object-detection      | yolo-v4-tf-416_INT8        | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h265-yolo-v4-tf-416                       | object-detection      | yolo-v4-tf-416_INT8        | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| od-h264-ssdlite-mobilenet-v2                 | object-detection      | ssdlite_mobilenet_v2       | mockrun    |
|                                              |                       |                            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+

Downloading an Object Detection Pipeline

Using the pipebench tool you can download the object detection pipeline and its accompanying models and sample media.

pipebench download od-h265-ssd-mobilenet-v1-coco

Expected Output

=========================================
Downloading od-h265-ssd-mobilenet-v1-coco
=========================================


===================
Downloading Runners
===================

Download dlstreamer
Building: ⠋
=================
Downloading Media
=================

Downloading: Pexels-Videos-1388365.temp
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 44.4M/44.4M [00:01<00:00, 41.0MiB/s]
        FFMPEG Converting
Converting: ⠙
==================
Downloading Models
==================

Download: ssd_mobilenet_v1_coco_INT8
Loading: ⠴
Pipeline Downloaded

Measuring Throughput

The following command reports total throughput running 1 stream (1 process, 1 total stream).

pipebench run od-h265-ssd-mobilenet-v1-coco --platform dgpu --streams 1 --media video/Pexels-Videos-1388365-h265 --duration 300

Expected Output

Pipeline:
	od-h265-ssd-mobilenet-v1-coco

 Runner:
	dlstreamer
 	dlstreamer.dgpu.runner-settings.yml

 Media:
	video/Pexels-Videos-1388365-h265

 Measurement:
	throughput
 	throughput.measurement-settings.yml

 Output Directory:
	/home/pipeline-zoo/workspace/od-h265-ssd-mobilenet-v1-coco/measurements/throughput/dlstreamer.dgpu/run-0001

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0000       0001     0.0000    0.0000    0.0000     0.0000
========================================================================

<SNIP>

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0001       0001  1369.1581 1369.1581 1369.1581  1369.1581
========================================================================

Pipeline                       Runner      Streams: 1
-----------------------------  ----------  -------------------------------------------------------------
od-h265-ssd-mobilenet-v1-coco  dlstreamer  Min: 1369.1581 Max: 1369.1581 Avg: 1369.1581 Total: 1369.1581

Measuring Stream Density

The following command measures stream density using 8 streams (1 process, 8 total streams).

timeout 300 pipebench run --measure density --platform dgpu od-h265-ssd-mobilenet-v1-coco --no-numactl --duration 500 --streams 8 --streams-per-process 8

Expected Output

 Pipeline:
	od-h265-ssd-mobilenet-v1-coco

 Runner:
	dlstreamer
 	dlstreamer.density.dgpu.runner-settings.yml

 Media:
	video/Pexels-Videos-1388365-h265

 Measurement:
	density
 	density.measurement-settings.yml

 Output Directory:
	/home/pipeline-zoo/workspace/od-h265-ssd-mobilenet-v1-coco/measurements/density/dlstreamer.dgpu/run-0004

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0000       0001     0.0000    0.0000    0.0000     0.0000
======================================================================== 

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0000       0001     0.0000    0.0000    0.0000     0.0000
======================================================================== 
<SNIP>
========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0008       0001   387.4471  393.1357  412.6040  3145.0855
========================================================================

Pipeline                       Runner      Streams: 8
-----------------------------  ----------  ----------------------------------------------------------
od-h265-ssd-mobilenet-v1-coco  dlstreamer  Min: 387.4471 Max: 412.6040 Avg: 393.1357 Total: 3145.0855

Object Classification Pipelines

stateDiagram
    direction LR
    state Object-Classification {
	direction LR
    state video_source {
	direction LR
		demux --> parse 
    }
	frames
	state classify {
	direction LR
		frames_1 --> crop_1
		objects_1 --> crop_1
		crop_1 --> scale_2
		scale_2 --> csc_2
		csc_2 --> inference_2
		inference_2 --> tensors_to_attributes_1
	}
    
    state detect {
	direction LR
		scale --> csc
		csc --> inference
		inference --> tensors_to_objects
    }
	
	state classify_N {
	direction LR
		frames_N --> crop_N
		objects_N --> crop_N
		crop_N --> scale_N
		scale_N --> csc_N
		csc_N --> inference_N
		inference_N --> tensors_to_attributes_N
	}
    state fork <<fork>>
	state join <<join>>
	objects-->fork
	frames-->fork
    media --> video_source
    video_source --> decode
    decode --> detect
    detect --> objects
	detect --> frames
	fork --> classify
	fork --> classify_N
	classify --> join 
	classify_N --> join 
	join --> objects_attributes
}
Loading 
+----------------------------------------------+-----------------------+----------------------------+------------+
| Pipeline                                     | Task                  | Models                     | Runners    |
+==============================================+=======================+============================+============+
| oc-h264-full_frame-efficientnet-b0           | object-classification | full_frame                 | mockrun    |
|                                              |                       | efficientnet-b0            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-ssdlite-mobilenet-v2-efficientnet-b0 | object-classification | ssdlite_mobilenet_v2       | mockrun    |
|                                              |                       | efficientnet-b0            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-yolov5s-416-efficientnet-b0          | object-classification | yolov5s-416                | mockrun    |
|                                              |                       | efficientnet-b0            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-ssdlite-mobilenet-v2-efficientnet-b0 | object-classification | ssdlite_mobilenet_v2       | mockrun    |
|                                              |                       | efficientnet-b0            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-full_frame-efficientnet-b0           | object-classification | full_frame                 | mockrun    |
|                                              |                       | efficientnet-b0            | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-yolo-v4-tf-resnet-50-tf              | object-classification | yolo-v4-tf                 | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-full_frame-resnet-50-tf              | object-classification | full_frame                 | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-full_frame-resnet-50-tf              | object-classification | full_frame                 | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-ssd-mobilenet-v1-coco-resnet-50-tf   | object-classification | ssd_mobilenet_v1_coco      | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-yolo-v4-tf-416-resnet-50-tf          | object-classification | yolo-v4-tf-416_INT8        | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-yolo-v4-tf-resnet-50-tf              | object-classification | yolo-v4-tf                 | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h264-yolo-v4-tf-416-resnet-50-tf          | object-classification | yolo-v4-tf-416_INT8        | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+
| oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf   | object-classification | ssd_mobilenet_v1_coco      | mockrun    |
|                                              |                       | resnet-50-tf               | dlstreamer |
+----------------------------------------------+-----------------------+----------------------------+------------+

Downloading an Object Classification Pipeline

Using the pipebench tool you can download an object classification pipeline and its accompanying models and sample media.

pipebench download oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf

Expected Output

======================================================
Downloading oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf
======================================================


===================
Downloading Runners
===================

Download mockrun
Download dlstreamer
Building: ⠋
=================
Downloading Media
=================

Downloading: Pexels-Videos-1388365.temp
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 44.4M/44.4M [00:00<00:00, 58.2MiB/s]
        FFMPEG Converting
Downloading: 20200711_dog_bark-1080p.temp
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 26.5M/26.5M [00:00<00:00, 46.6MiB/s]
        FFMPEG Converting
Converting: ⠹
==================
Downloading Models
==================

Download: resnet-50-tf_INT8
Download: resnet-50-tf
Download: ssd_mobilenet_v1_coco_INT8
Download: ssd_mobilenet_v1_coco
Loading: ⠦
Pipeline Downloaded

Measuring Throughput

The following command reports total throughput running 1 stream (1 process, 1 total stream).

pipebench run oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf --platform dgpu --streams 1 --media video/20200711_dog_bark-1080p-h265

Expected Output

Pipeline:
        oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf

 Runner:
        dlstreamer
        dlstreamer.dgpu.runner-settings.yml

 Media:
        video/20200711_dog_bark-1080p-h265

 Measurement:
        throughput
        throughput.measurement-settings.yml

 Output Directory:
        /home/pipeline-zoo/workspace/oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf/measurements/throughput/dlstreamer.dgpu/run-0000

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0000       0001     0.0000    0.0000    0.0000     0.0000
========================================================================
<SNIP>
========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0001       0001  1666.5071 1666.5071 1666.5071  1666.5071
========================================================================

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0001       0001  1689.3651 1689.3651 1689.3651  1689.3651
========================================================================

Pipeline                                    Runner      Streams: 1
------------------------------------------  ----------  -------------------------------------------------------------
oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf  dlstreamer  Min: 1689.3651 Max: 1689.3651 Avg: 1689.3651 Total: 1689.3651

Measuring Stream Density

The following command measures stream density using 8 streams (1 process, 8 total streams).

timeout 300 pipebench run --measure density --platform dgpu oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf --media video/20200711_dog_bark-1080p-h265 --no-numactl --duration 240 --streams 8 --streams-per-process 8

Expected Output

Pipeline:
        oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf

 Runner:
        dlstreamer
        dlstreamer.density.dgpu.runner-settings.yml

 Media:
        video/20200711_dog_bark-1080p-h265

 Measurement:
        density
        density.measurement-settings.yml

 Output Directory:
        /home/pipeline-zoo/workspace/oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf/measurements/density/dlstreamer.dgpu/run-0001

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0000       0001     0.0000    0.0000    0.0000     0.0000
========================================================================
<SNIP>
========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0008       0001   296.7597  304.2268  316.2266  2433.8145
========================================================================

========================================================================
Iteration   Streams  Processes    Minimum   Average   Maximum      Total
========================================================================
     0000      0008       0001   302.6452  305.9524  316.2266  2447.6194
========================================================================

Pipeline                                    Runner      Streams: 8
------------------------------------------  ----------  ----------------------------------------------------------
oc-h265-ssd-mobilenet-v1-coco-resnet-50-tf  dlstreamer  Min: 302.6452 Max: 316.2266 Avg: 305.9524 Total: 2447.6194