@jamesh I don't think you are answering my point.
Right now, when I type rpicam-hello, it seems to me that PISP is only used for stats to calculate some averages and then allow to tune exposure and gain. But any sensor already provides averages. Perhaps PISP can do a better job but the improvement seems quasi insignifiant.
The first document you are mentioning talk about camera tuning, exactly my point above. It doesn't talk about color conversion or scaling unless I'm mistaken.
When you have a 4k image, the heavy operation is image color conversion and scaling (and encoding).
And BTW, Raspberry PI 5 already doesn't have a hardware encoder!!! (removed from Pi 4). So when I do rpicam-vid all the hard work is done by Neon: color conversion, scaling and encoding. That takes all CPU.
I'm asking again my question differently: how can I do at least color conversion and scaling in hardware? Is there an example somewhere using any library that does color conversion in hardware????
Even if you are using libcamera and you are forcing an output that makes sense (like scaling / 2 and I420), the library refuses the configuration. And that's exactly what is happening in rpicam-*. It gets out from the PISP the same format and resolution as the sensor (except _8 -> _16 which is worse as it doubles memory usage and parsing time).
Right now, when I type rpicam-hello, it seems to me that PISP is only used for stats to calculate some averages and then allow to tune exposure and gain. But any sensor already provides averages. Perhaps PISP can do a better job but the improvement seems quasi insignifiant.
The first document you are mentioning talk about camera tuning, exactly my point above. It doesn't talk about color conversion or scaling unless I'm mistaken.
When you have a 4k image, the heavy operation is image color conversion and scaling (and encoding).
And BTW, Raspberry PI 5 already doesn't have a hardware encoder!!! (removed from Pi 4). So when I do rpicam-vid all the hard work is done by Neon: color conversion, scaling and encoding. That takes all CPU.
I'm asking again my question differently: how can I do at least color conversion and scaling in hardware? Is there an example somewhere using any library that does color conversion in hardware????
Even if you are using libcamera and you are forcing an output that makes sense (like scaling / 2 and I420), the library refuses the configuration. And that's exactly what is happening in rpicam-*. It gets out from the PISP the same format and resolution as the sensor (except _8 -> _16 which is worse as it doubles memory usage and parsing time).
Statistics: Posted by gregoiregentil — Tue Nov 12, 2024 7:38 am