Changes related to PR in backend common utility. @tanmayv25 I notice the backend itself doesn't use BackendInputCollector for input collection, is there a reason for doing so, or I should change it to use BackendInputCollector if the input is not contiguous
Tested on resnet50_fp16 and checked that the change does avoid unnecessary copy as compute input time is roughly the same before / after the change (~20 usec). Compute input with copy will be ~90 usec.
Changes related to PR in backend common utility. @tanmayv25 I notice the backend itself doesn't use
BackendInputCollectorfor input collection, is there a reason for doing so, or I should change it to useBackendInputCollectorif the input is not contiguousTested on resnet50_fp16 and checked that the change does avoid unnecessary copy as compute input time is roughly the same before / after the change (~20 usec). Compute input with copy will be ~90 usec.