Hi,
I have two dlc models, that contain only one layer - instancenorm2d.
Input sizes are 1,224,224,64 and 1,224,224,65.
Both of them work fine on the CPU (Snapdragon 855). On the GPU the first one works, but the second crashes with the following error:
2019-05-28 07:16:16,432 - ERROR - snpe_bench: ['-------------------------------------------------------------------------------', 'Model String: N/A', 'SNPE v1.25.1.310', '-------------------------------------------------------------------------------', 'Processing DNN input(s):', 'random_inputs/random_input_1.raw', 'error_code=808; error_message=OpenCL check has failed. OpenCL Error (-54) CL_INVALID_WORK_GROUP_SIZE; error_component=GPU Runtime; line_no=131; thread_id=-223193944; opencl_error=-54']
2019-05-28 07:16:16,432 - ERROR - snpe_bench: ['-------------------------------------------------------------------------------', 'Model String: N/A', 'SNPE v1.25.1.310', '-------------------------------------------------------------------------------', 'Processing DNN input(s):', 'random_inputs/random_input_1.raw', 'error_code=808; error_message=OpenCL check has failed. OpenCL Error (-54) CL_INVALID_WORK_GROUP_SIZE; error_component=GPU Runtime; line_no=131; thread_id=-223193944; opencl_error=-54']
2019-05-28 07:16:16,433 - WARNING - snpe_bench: Failed to perform benchmark for GPU_timing.
Looks like some problem with local and global worksizes in OpenCL implementation. Maybe they are set equal, and on channelsize=65 it just exceeds the limits of OpenCL local workgroup size for this hardware.
I can send models for you to test if you provide some email.
Best Regards,
Aleksei Ivakhnenko.