I know there's at least one anon here who writes CUDA accellerated algorithms.
How do I get into it.
Hard mode: Same question, but without an NVIDIA card.
>>61182171
webgl/2 has vertex and color shaders, running on an apu here =)
>>61182171
OpenCL
>>61182171
You cant run cuda programs on non nvidia cards.
You have to use OpenCL which is very similar to cuda.
Here is a list of books
https://www.khronos.org/developers/books/
It depends on your algorithm. You can think of graphics cards as highly parallel calculators. Its best when you have lots of small independent operations. For example matrix multiplication.
If your problem is something like this Google/Stackoverflow are pretty much all you need though. Lots of example code out there.
Do you really need raw CUDA?
Try more specialized approach if you can, like Theano, Keras, Tensorflow and so on.
Raw CUDA is pretty hard, so expect your shit to run much slower because you won't know how to optimize it.
http://webglreport.com/?v=2
>>61182438
You're an SJW cuck, aren't you?
>>61182480
TensorFlow uses CUDA through cuDNN.
>>61182535
But it doesn't use it directly and thus doesn't require optimizing CUDA code by hand, resulting in much better performance if you aren't a CUDA buddhajesus.
>>61182399
>>61182438
>>61182480
Thanks for the truly high quality replies anons. I'm not even being sarcastic. Will be lurking this thread, and probably coming back for more OpenCL discussion at a later time.
>>61182547
I agree. It's better to use nVidia libraries for CUDA instead of writing kernels by hand unless you're very experienced with numerical analysis and the underlying hardware idiosyncrasies, they usually handle the hardware more smart. But it won't be redundant to have a notion how it works under the hood.
>>61182171
I started learning CUDA this week, so far I've wrote a matrix multiplier and a kernel convolver. Its hard but managable.
>>61182171
I tried setting it up on fedora and it's a pain in the ass
More or less on any other distro except ubuntu 16.04/14.04
>>61183172
Debian 9 is OK.
>>61182171
The udacity course is a good place to start
>>61183060
Fuck you nigger, you're the chicken
>>61183280
Pajeet detected
>>61182171
If you don't have an nvidia GPU you can always spin up an AWS with one and use it to experiment. I guess you could learn OpenCL as well but CUDA will continue to destroy it in performance so long as nvidia wants it to.
>>61182480
> using someone elses framework
LMAO fucking nigger.