search

quic / qidk

Other
76 stars 16 forks source link

Implement cropscale and normalization Via HVX #12

Closed PureHing closed 3 months ago

PureHing commented 4 months ago

@quic-rneti @quic-shubpate @quic-ppant Hi,

  1. I got NV12 image data from the camera pipline of sm8550. I want to send the NV12 data directly to HTP for detection model inference. The input of the detect model is RGB and normalized, so for my NV12 data I have to execute Cropscale and normalization. In my application scenario, the GPU resources are not enough and can no longer be used for model pre-processing. I would like to ask if there is any good solution? My model input is 960x448 and only requires 2-3ms of inference in htp. If I want to use hvx_cv to implement cropscale and normalization, assuming my NV12 is 4K or larger, what kind of performance can I get. Refer to this link :https://github.com/quic/qidk/blob/10fc5c9bbf918163aa54087a4eb51ff4be1ed9a8/QHCI/hvx_cv/src/rgb2yuv/dsp/rgb2yuv_imp.c#L2663

2.BTW,I have tested that the performance of the ds module of evs is satisfactory, but resize with crop parameters cannot be implemented. Is it possible to support it?

can you give me some advice?

quic-rneti commented 3 months ago

We are discussing this issue in e-mail. So closing the github issue.

PureHing commented 3 months ago

We are discussing this issue in e-mail. So closing the github issue.

@quic-rneti Sorry to bother you, I have replied to your email on March 12th, but it seems that you have not replied to me.