the current qnn sdk i install from qpm , it's run a resnet18 model with htp backend using int8 precision is ok but using fp16 is not work ,result is empty . we get aptiv support and use a auto version qnn sdk which get from qualcomm to inferrence this model using fp16 it's work. but aptiv has no right to delivery this auto version qnn sdk to us ,it's need qualcomm supports
QNN cannot run fp16 htp inferrence on aptiv 8295
Posted: Mon, 2023-07-03 00:55
Dear developer,
Pls try to analyze the cDSP logs for your issues.
BR.
Wei