search

nvdla / sw

NVDLA SW
Other
489 stars 193 forks source link

nvdla_runtime hanging with Resnet-50 loadable in virtual platform! #165

Open AmritaMathuriya opened 5 years ago

AmritaMathuriya commented 5 years ago

Hi everyone, I tried compiling and executing the ResNet-50 workload with the provided prebuilt binaries with the virtual platform. I am using the virtual platform from the Nvidia provided NVDLA AMI nvdla_vp_fpga_ami_ubuntu.

I used the following commands: ./nvdla_compiler --cprecision fp16 --configtarget nv_full --prototxt ../runBenchmarks/ResNet-50-deploy.prototxt --caffemodel ../runBenchmarks/ResNet-50-model.caffemodel

./nvdla_runtime --loadable fast-math.nvdla --image new_dog.jpg creating new runtime context... [ 78.600610] random: crng init done

The emulator is hanging at that point. Could anyone please provide helpful pointers? If ResNet is too big, is there a sample case I could try to test my setup?

neroangelo296 commented 5 years ago

Hi @AmritaMathuriya, I seem to be stuck at the compiler step as well. However, I have not been able to generate the .nvdla file. For some reason every-time I compile, the wisdom.dir directory is generated without any output file.

Do you have any ideas? Could you maybe also send your .prototxt and .caffemodel to make sure we're on the same page.

Awaiting your reply.

AmritaMathuriya commented 5 years ago

@neroangelo296 The .nvdla file is generated in the same directory and not inside the wisdom.dir. I used the prototxtand caffemodel from the same link as suggested in this repo, for the Resnet-50.

https://github.com/KaimingHe/deep-residual-networks The above link has a link to the models.

https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777

neroangelo296 commented 5 years ago

Thanks, I was able to produce the .nvdla files correctly.

Also, are you running the runtime environment on Arm or x86? As far as I know, the runtime provided is compiled only for Arm architectures.

AmritaMathuriya commented 5 years ago

The binaries that are hanging are pre-compiled nvdla compile and run time for ARM architecture. I tried running those in the provided virtual platform. I also have another setup where I have nvdla_runtime binary for the risc-v setup, but I am getting a different error there.

neroangelo296 commented 5 years ago

Could you post the error log for the RISC setup? Also, what are you running the RISC-V core out of? An FPGA?

AmritaMathuriya commented 5 years ago

Please check this issue. Yes, I am running the RISC-V core on an FPGA. https://github.com/nvdla/sw/issues/164