As the unsupported_ops and op_sets_to_ignore maps are defined with worker_id, not device id, I think using tmp_worker_id corresponding with device_flag is not correct.
Suppose,
Worker 0: CPU with CPUMaskFlag::kBig.
Worker 1: CPU with CPUMaskFlag::kLittle,
Worker 2: GPU with CPUMaskFlag::kAll
case.
In that manner, we should find the corresponding unsupported ops' information using worker_id, not device_flag.
In
band/model_analyzer.cc
, I have a question for this code block. https://github.com/mrsnu/band/blob/236378ef119d8b744b39932ea7b8a3481e0c337b/band/model_analyzer.cc#L354-L362As the
unsupported_ops
andop_sets_to_ignore
maps are defined with worker_id, not device id, I think usingtmp_worker_id
corresponding withdevice_flag
is not correct. Suppose, Worker 0: CPU with CPUMaskFlag::kBig. Worker 1: CPU with CPUMaskFlag::kLittle, Worker 2: GPU with CPUMaskFlag::kAll case. In that manner, we should find the corresponding unsupported ops' information usingworker_id
, notdevice_flag
.My suggestion is changing the code into
Could you suggest any ideas?