@halirutan I enjoy lots of kinds of cuisines, (some of) Chinese, Japanese, Italian, Thai, Indian, ... But I myself is really a bad cook. It was usually my other half doing the cooking and me enjoying the exercise of dish-washing. I do know how to bake bread and cake though! -- but that's too cumbersome :)
@kirma During the early stage in Wuhan, we lost several senior professors there. I'm afraid I'm not optimistic about the upcoming situation in Princeton U or any other universities at the epicenter..
@LukasLang & others, is there a QA that might be useful in understanding how to write code which uses multithreading? Maybe another resource? (My CS & coding skills are all self-taught, you see.)
I think each kernel is its own process, not threading.
My poor-girl's multithreading on GPU is by abusing the neural network framework. Most numerical computing can be ported that way, avoiding coding in CUDA directly.
Yeah that definitely makes sense, @Silvia —once I considered it on that level. I see examples from Henrik & others wherein a semi-complicated matrix with some semi-regular structure can be computed in microseconds, and I presume this mainly to be related to the ability to compile into c—in general I wonder how you can take some matrix of fixed dimension and compute it as quick as possible. All numeric in the end, for an Eigensystem call, of course. Re: GPU multithreading OH MY what!
@Silvia One’s boring is another’s Sunday Funday ;) thank you for the tip!! I thought memoization was the key, for example having some rate term being changed while you keep the angles constant, so you might as well remember the computations which only use the angles, but (understandably so) if you wrap in module or block to keep local contexts local, that doesn’t help, logically because you’ve(I’ve) locked the computations away under the protection of the local context.
And I understand all of that, now, but I have no clue where to go from there. Something something lack of advanced/basic linear algebra knowledge something something basics of coding constructs :s
@CATrevillian I tried "programming" in neural network way, it felt more like programming in assembly, no local context, no scopes -- though scope can be emulated by hierarchy NetGraph.
The result is a normal NetGraph / NetChain, so it should be fine to plug and play in NetTrain or NetGradient.
