DLA supports only 3 subgraphs per DLA core

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
1

I want to use DLA core to accelerate the inference on my xavier. So when I created a conv layer, I used nvinfer1::Builder::setDeviceType() to let this layer run on DLA core. But when I created the Engine, it showed me this message:
…
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_103
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_104
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_105
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_82
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_94
WARNING: DLA supports only 3 subgraphs per DLA core. Switching to GPU for layer conv_106
…
Does that mean only 3 conv layers can run on a DLA core so that other layers can only run on GPU?

2

Hi,

Current max DLA subgraph is limited to 3.
The value will increase to 32 after our efficient memory partition implemented, which should be DLASW 1.2.

Thanks.

3

Is this feature still under development? Just wondering since the DLA SW repo has been stale since August 2018.

4

Hi,

DLA 1.2 will be included in our next release. (JetPack4.2.1)
Please pay attention to our announcements.

Thanks.

5

Hi,

I am using Jetpack 4.3 version and getting error for 8, not 3 or 32. This affects 95 layers out of 161 layers. I checked the updates but could not find a solution for this. I am using yolov4 config. This switches are on shortcut and convolution layers.

WARNING: DLA supports only 8 subgraphs per DLA core. Switching to GPU for layer XXX

Do you have any idea/suggestion on this?