Thank you! I was coming toward implementing the latter “if drag/release don’t reset” at least to try it out. And thank you for these posts in the first place. They give really great insight on how to do “immediate mode” interactivity.

LikeLike

You are right. The code I posted is a simplification, extracted from a more complex app.
In the simplification process I may have broken something.
In the app this code is different:
– gesture retains the selected code and give access to it through a function int getCurrentSelectionCode()
– gesture pick returns a bool instead of the picking code. False if the picking is Graphics::k_noSelectionCode with early return if clickEnbaled || clickDrag.
By early returning the function will not update its last selection code during drag or button release events. In early returns it
returns retained code != Graphics::k_noSelectionCode.
– the implementation use the return value true to call:
int selectionCode = gesture.getCurrentSelectionCode();
forEachTool([&](ManipulationTool* tool) { tool->setActiveCode(selectionCode); });

This has the effect of restoring the selection code of the manipulator. I am sure it can be done in other ways, such as “if drag/release don’t call reset…
I’ll try to amend the code in the blog post.

LikeLiked by 1 person

LikeLiked by 1 person

Indeed I understand that you didn’t state it clean as that, but it’s implied that NCHW is faster (cause it faster in SIMD cpu? case)

“SIMD instructions prefer reading SOA data because, with each individual load and store instructions, the processor can fill wide registers with several elements accessed consecutively and sequentially. This makes good use of the memory latency and bandwidth, resulting in faster processing speeds.”

And it’s not true for neither NVidia nor Apple or even Intel on each of this GPU’s channel last is faster and preferred for both FP32/FP16 and BF16/TF32 precision (if available). [I’m not sure about AMD as I never worked with it].

LikeLiked by 1 person

Sure, that’s understandable. Good luck and looking forward to the next chapter.

LikeLike

Thank you for the comment. It is correct, NVIDIA GPUs may favor other data layout, and that depends on the precision required and the instruction set available.
In this post I am not looking at performance yet, instead I am showing that there are many options to consider, and I am trying to keep the text as simple as I can for the purpose of the narration. How the choice of tensor layout, data precision and the hardware come together is something I will write about in a future post.

LikeLike

Hi Jakub, life got in the way. I will resume the series in the next couple of months. Thank you for the patience.

LikeLike

LikeLiked by 1 person

LikeLike

LikeLike