Deepstream-app play 8 stream input video is slow

[terry.huang]

Hi

I use deepstream-app to play 8 stream input video.

deepstream-app -c source2_1080p_dec_infer-resnet_demux_int8.txt

Additional, I enable DLA by adding the below line in config_infer_primary.txt.
enable-dla=1
use-dla-core=0

But VLC shows the RTSP video is very slow.
I also use tegrastats to check the DLA is disabled, even enable-dla=1.

How to make it faster ?
My environment: JP6, Orin NX 16G

tegrastats

08-21-2024 14:33:46 RAM 5974/15656MB (lfb 29x2MB) SWAP 839/7828MB (cached 4MB) CPU [27%@729,11%@729,8%@1121,35%@1984,15%@729,15%@729,15%@729,14%@729] EMC_FREQ 12%@2133 GR3D_FREQ 27%@[305] NVENC 10%@115 NVDEC 34%@230 NVJPG off NVJPG1 off VIC 24%@256 OFA off NVDLA0 off NVDLA1 off PVA0_FREQ off APE 174 cv0@41.406C cpu@43.437C soc2@40.812C soc0@41.906C cv1@42.656C gpu@40.406C tj@45.156C soc1@45.156C cv2@41.375C
08-21-2024 14:33:47 RAM 5974/15656MB (lfb 29x2MB) SWAP 839/7828MB (cached 4MB) CPU [25%@1984,14%@1984,6%@1984,38%@1984,14%@729,12%@729,13%@729,8%@729] EMC_FREQ 12%@2133 GR3D_FREQ 61%@[305] NVENC 40%@115 NVDEC 44%@115 NVJPG off NVJPG1 off VIC 50%@268 OFA off NVDLA0 off NVDLA1 off PVA0_FREQ off APE 174 cv0@41.406C cpu@43.593C soc2@40.875C soc0@41.937C cv1@42.593C gpu@40.281C tj@44.937C soc1@44.937C cv2@41.343C

Thanks

[fanzh]

could you share the whole log? wondering if the DLA engine is created and loaded.

[terry.huang]

deepstream-app.log (36.8 KB)

The log is attached.
DLA0 still is disabled, even enable-dla=1 and use-dla-core=0 in config_infer_primary.txt.
NVDLA0 off NVDLA1 off

[terry.huang]

I found the video player still is slow, even osd is disabled. (see attachment)
if the number of video stream is down to 4, the video showing will back to normal speed.

[fanzh]

from the log " deserialized trt engine from :/opt/nvidia/deepstream/deepstream-7.0/samples/configs/deepstream-app/…/…/models/Primary_Detector/resnet18_trafficcamnet.etlt_b2_gpu0_int8.engine", the app loaded the GPU engine not DLA engine. please set model-engine-file to the DLA engine.
if still can’t work, please share the config_infer_primary.txt.

[terry.huang]

config_infer_primary.txt (3.8 KB)
deepstream-app-dla.log (117.3 KB)

Hi

It is not improved if DLA0 is enabled and motion is still very slow in video.

==> From log, It seems many layers are unsupported on DLA.
WARNING: [TRT]: Layer ‘block_1a_conv_1/kernel’ (CONSTANT): Unsupported on DLA. Switching this layer’s device type to GPU.
WARNING: [TRT]: Layer ‘block_1a_conv_1/bias’ (CONSTANT): Unsupported on DLA. Switching this layer’s device type to GPU.
WARNING: [TRT]: Layer ‘block_1a_bn_1/moving_variance’ (CONSTANT): Unsupported on DLA. Switching this layer’s device type to GPU.

==> DLA is usually idle.
08-22-2024 10:06:25 RAM 6540/15656MB (lfb 17x2MB) SWAP 823/7828MB (cached 4MB) CPU [16%@1984,6%@1984,2%@1984,7%@1984,2%@1984,3%@1984,3%@1984,0%@1984] EMC_FREQ 7%@3199 GR3D_FREQ 4%@[917] NVENC 18%@115 NVDEC 57%@166 NVJPG off NVJPG1 off VIC 27%@256 OFA off NVDLA0 0%@614 NVDLA1 off PVA0_FREQ off APE 174 cv0@47.218C cpu@49.125C soc2@46.375C soc0@47.125C cv1@48.093C gpu@45.625C tj@50.656C soc1@50.687C cv2@46.468C
08-22-2024 10:06:26 RAM 6541/15656MB (lfb 17x2MB) SWAP 823/7828MB (cached 4MB) CPU [16%@1984,5%@1984,5%@1984,7%@1984,6%@1984,2%@1984,3%@1984,3%@1984] EMC_FREQ 8%@3199 GR3D_FREQ 4%@[917] NVENC 7%@115 NVDEC 48%@179 NVJPG off NVJPG1 off VIC 47%@217 OFA off NVDLA0 0%@614 NVDLA1 off PVA0_FREQ off APE 174 cv0@47.281C cpu@49.031C soc2@46.531C soc0@47.187C cv1@48.25C gpu@45.625C tj@50.687C soc1@50.687C cv2@46.75C
08-22-2024 10:06:27 RAM 6541/15656MB (lfb 17x2MB) SWAP 823/7828MB (cached 4MB) CPU [13%@1984,3%@1984,4%@1984,2%@1984,4%@1984,4%@1984,6%@1984,5%@1984] EMC_FREQ 7%@3199 GR3D_FREQ 0%@[917] NVENC 31%@115 NVDEC 54%@115 NVJPG off NVJPG1 off VIC 63%@281 OFA off NVDLA0 0%@614 NVDLA1 off PVA0_FREQ off APE 174 cv0@47.156C cpu@49.093C soc2@46.468C soc0@47.093C cv1@48.218C gpu@45.593C tj@50.718C soc1@50.718C cv2@46.531C
08-22-2024 10:06:28 RAM 6542/15656MB (lfb 17x2MB) SWAP 823/7828MB (cached 4MB) CPU [16%@1984,2%@1984,5%@1984,3%@1984,8%@1984,5%@1984,5%@1984,2%@1984] EMC_FREQ 7%@3199 GR3D_FREQ 0%@[917] NVENC 14%@115 NVDEC 75%@115 NVJPG off NVJPG1 off VIC 40%@268 OFA off NVDLA0 42%@614 NVDLA1 off PVA0_FREQ off APE 174 cv0@47.5C cpu@49.062C soc2@46.531C soc0@47.281C cv1@48.125C gpu@46.093C tj@50.687C soc1@50.687C cv2@46.718C

Thanks

[fanzh]

There are eight sources. noticing the app load bath-size engine, please set batch-size=8 in nvinfer’s configuration file, and set batch-size=8 in [streammux] part of deepstream-app cfg, and set sync=0 in [sink] of deepstream-app cfg.

[terry.huang]

Hi

The change can speedup the motion in display but this display is not very clear.
I change width=1920 and height=1080 in [tiled-display], but it still is same.
Hot to improve it ?

Thanks

[terry.huang]

Hi

I fix it by changing bitrate.

Thanks for support