Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Andrey_S_Intel1
Employee
99 Views

Kind of Unet architecture works fine on CPU but generate corrupted result on GPU

Kind of Unet architecture works fine on CPU but generate corrupted result on GPU.

We have a kind of Unet which works fine on both devices (CPU and GPU), but when we reduce the number of generated convolutions in each layer by some constant factor, the GPU inference stop generate valid result (start to see garbage like vertical stripes instead of image). There is no errors or worming during model generation or inference, I am wandering if there is some memory alignment related to openCL and GPU that cause to this issue and does not taken into account by open vino? 

 

 

0 Kudos
5 Replies
om77
New Contributor I
99 Views

Hi Andrey,

just wanted to ask you, do you see the "same" Unet output in GPU FP32 and GPU FP16 modes?

My issue with Unet is that GPU in FP16 mode output is incorrect while GPU in FP32 mode is ok.

And MYRIAD FP16 output is ok.

nikos1
Valued Contributor I
99 Views

Do you get GPU issues with both FP16 and FP32 IR?

Andrey_S_Intel1
Employee
99 Views

yes, the issue appear with both FP16/32 IRs.

nikos1
Valued Contributor I
99 Views

Interesting!  So FP32 CPU is good but FP32 GPU has incorrect results.

Any way to repro this?

BTW what is the GPU (or CPU) model?

Andrey_S_Intel1
Employee
99 Views

Hi Nikos,

 

I can't share the code outside of Intel, hope that some one from Intel support internally will assist.

Reply