01-26-2021 08:26 PM
I see no mention of set_window_size() in the latest UG1076. Has support for dynamic window sizing been removed?
02-12-2021 08:19 AM
HI @e_ensafi
As you have a patch for the asynchronous windows using templates and I clarified about the set_window_size() could kindly you mark this topic as solved?
Thanks
01-27-2021 01:32 AM
Hi @e_ensafi
The set_window_size() has many limitation and was mostly misused. For this reason, it was removed from the UG1076.
We are currently in the process of deciding if this should be kept (with all the limitation mentioned in the UG1076) or if it should be completely removed.
This is important to note that this API is not really changing the size of the window but only the size the kernel is perceiving. So in the graph the maximum window size had to be defined and the DMA transfer are happening on the full window size. I never try what it does for the input data but I believe this would discard some of the input data.
I found that using the asynchronous windows to dynamically change the amount of data to read was a much better solution. At minima, this avoid inefficient DMA transfers.
What is exactly your use case on this?
01-27-2021 08:30 AM
@florentw Interesting that you should mention asynchronous windows, because the only reason why I revisited dynamic window sizes was the fact that when templates are mixed with asynchronous windows, compilation fails (as reported in another thread). This was a showstopper for us, so I had to look into other options.
01-27-2021 09:14 AM
Hi @e_ensafi
I guess the option for you for the moment is to avoid using templetized kernel when using asynchronous window.
I let the development know that the issue with asynchronous windows was a high priority. I am waiting to hear from them.
02-12-2021 08:19 AM
HI @e_ensafi
As you have a patch for the asynchronous windows using templates and I clarified about the set_window_size() could kindly you mark this topic as solved?
Thanks