hi, we tried to run vgg-ssd-caffe on snapdragon 845. We converted the dlc with sdk 1.24 on Ubuntu 16.04.
The dlc could be loaded successfully on cpu runtime, but it failed on gpu runtime.
Then we tried the alexnet, it cound be loaded successfully on both cpu and gpu runtime.
In fact, we has ever run vgg-ssd on gpu successfully, but the dlc was converted by sdk 1.15.
So, is there any fatal change between 1.15 and 1.24(we has set batch to 1)? Same prototxt, same caffemodel, dlc converted by 1.15 could run on gpu, dlc converted by 1.24 could not.
Please help!
Please check the layers mentioned in SNPE Docs, The supported layers will only run on the GPU runtime, else it will fallback to CPU runtime.
Please cross-verify the layers of vgg-ssd-caffe with the supported layer name.
Thanks for your promptly reply!
I have checked the Doc, and I confirm all the layers can run on GPU runtime.
In fact, we have ever successfully run vgg-ssd-caffe on GPU with SDK 1.15. However now we try to update the SDK to 1.24, same .prototxt, same .caffemodel and same terminal commands, but the new .dlc cannot be loaded on gpu runtime TAT.
So I guess there may be some important updates between the two SDK version.
Any ideas please