[Leo Fang, chainer] @ @Emilio @ @ @beam2d I’ve seen this message appearing after every PR merged to conda-forge/cupy-feedstock. I think it’s because we added you as maintainers but you did not accept the invitation. If you would like to be listed as a maintainer, could you please accept the invitation? If you prefer not to be listed, could you let me know so that I can take your ID out and suppress the bot
Hi, I'm looking to work on ILU cupy/cupy#2749 - I wrote iterative solver kernels in CUDA before, including ILU.
Before I start, I have some questions:
- How does cupy's architecture support functions calling on single and multiple gpus? How are number of threads and number of blocks defined by cupy in each function?
- Sometimes there are NVIDIA functions (like cublas, cusparse) but sometimes they need to be written by developers. In what kind of situations do the cupy library developers write custom kernels to support new functions?
Nonetheless as long as there are cublas and cusparse functions in place, the ILU kernel can be written using those high level APIs. So my questions are mostly for the general case. Thank you.
[Akifumi Imanishi, chainer] @wongalvis14
- Most of CuPy's function has a highly compatible interface with NumPy, so it does not have the interface to specify the block size and the grid size.
- The feature of the custom kernels is documented here: https://docs-cupy.chainer.org/en/stable/tutorial/kernel.html. It can specify the block size.
[Emilio, chainer] > How are number of threads and number of blocks defined by cupy in each function?
Right now almost all cupy functions launch kernels with the same number of threads per block. the number of blocks is defined by the size of the arrays.
We are working on some kind of auto-tuning for these parameters .
Right now almost all cupy functions launch kernels with the same number of threads per block. the number of blocks is defined by the size of the arrays.
We are working on some kind of auto-tuning for these parameters .
[nishino, chainer] To complement the above comment, this is the still-WIP PR for the auto-tuning feature: cupy/cupy#2731
Even lower-level class
Even lower-level class
cupy.cuda.Function has the interface to specify threads/blocks but its mostly used internally.
[nishino, chainer] > How does cupy's architecture support functions calling on single and multiple gpus?
You can swith the current device by using
You can swith the current device by using
cupy.cuda.Device:with cupy.cuda.Device(1):
a = cupy.exp(cupy.ones((2, 3), 'f'))
print(a.device) # Prints <CUDA Device 1>
Hi, I found a function with a type of syntax I've never seen before https://github.com/cupy/cupy/blob/9f9ef7d15d632619f650471e9407722ce6a73438/cupy/_sorting/count.py#L27
<Sameroom> Whoops! You've exceeded your daily message limit on this Sameroom account (it will reset in 5 hours and 10 minutes). If you have too many Tubes for your budget, pause or delete some Tubes on the Manage page. If you dont have a subscription, visit pricing to upgrade to unlimited messaging.
<Sameroom> Whoops! You've exceeded your daily message limit on this Sameroom account (it will reset in 5 hours and 10 minutes). If you have too many Tubes for your budget, pause or delete some Tubes on the Manage page. If you dont have a subscription, visit pricing to upgrade to unlimited messaging.
Hi, I've written a draft for Cupy sparse mean #2674 https://gist.github.com/wongalvis14/e4c1af431c20cb18d5457a372a32c3ec
It's not optimized yet in comparison to numpy, and cupy implementations that I have yet to observe
Please give me some feedback on how to keep it up to production quality, thanks
It's not optimized yet in comparison to numpy, and cupy implementations that I have yet to observe
Please give me some feedback on how to keep it up to production quality, thanks
That's one thing I notice from scipy's implementation. I'm wondering if there's code I can refer to within cupy for cupy-oriented optimizations.
Also I don't know if it is correct to handle the row sum case (axis=0) by returning a flattened array, I just did what was most simple and intuitive to match the output formats with scipy and cupy dense mean.
It seems that dividing first then sum them up would require elementwise division, which isn't implemeneted. I also see that simpler elementwise operations such as addition isn't implemented?https://github.com/cupy/cupy/blob/ebe2cae7b2a6efb2b601d7da1a6471eeb0434a6e/cupyx/scipy/sparse/compressed.py#L426
Adding or dividing with a scalar shouldn't have many tricky cases, is there already a sparse elementwise kernel in cupy?
I have improved it to reduce the risk of precision error https://gist.github.com/wongalvis14/91400ced33778fd24e315ed85f5c6655
Oh actually I should put it under cupyx/scipy/sparse/data.py alongside other elementwise functions like power right?
<Sameroom> Whoops! You've exceeded your daily message limit on this Sameroom account (it will reset in 24 hours and 0 minutes). If you have too many Tubes for your budget, pause or delete some Tubes on the Manage page. If you dont have a subscription, visit pricing to upgrade to unlimited messaging.
[Emilio, chainer] base is fine as sum is defined there
<Sameroom> Whoops! You've exceeded your daily message limit on this Sameroom account (it will reset in 23 hours and 51 minutes). If you have too many Tubes for your budget, pause or delete some Tubes on the Manage page. If you dont have a subscription, visit pricing to upgrade to unlimited messaging.
I saw that one of the projects is to support rng https://github.com/cupy/cupy/wiki/GSoC-2020-Project-Ideas
cuRand is imported into cupy and cupy.random already exists, right? What is left to be done for that project?
cuRand is imported into cupy and cupy.random already exists, right? What is left to be done for that project?
Hey everybody! I am Rushabh Vasani, A 2nd-year I.T undergrad from India. I am highly interested in working with cupy this summer under GSoC. I am interested in working on "CuPy coverage of NumPy functions". And I even have contributed to cupy previously. And I really looking forward to contributing more. Right now I am listing up the numpy functions which are not there in cupy for my reference. And trying to raise some more PRs in the project. Can you guys guide me further to get started with this project and the community? Thanks in advance!
[Leo Fang, chainer] Quick question for the merged PR 2584: I thought the preference for memoization is to use
cupy.util.memoize, but now both ufuncs/element-wise kernels (never use it before) and reduction kernels (changed after 2584) move away from it? Is there a performance reason, or just to avoid decorators in Cython codes?
[Emilio, chainer] you should ask imanishi
[Kenichi Maehashi, chainer] I’ve updated development policies page to covers topics often pointed out during the pull-request review. https://github.com/cupy/cupy/wiki/Development-policies
[Leo Fang, chainer] I saw a folk from Autograd made a fork to support CuPy, but it was never merged into Autograd:
https://github.com/ericmjl/autograd-cupy
I am curious its performance and extensibility compared to Chainer’s (if there’s one). Also, it might be worth adding to the
https://github.com/ericmjl/autograd-cupy
I am curious its performance and extensibility compared to Chainer’s (if there’s one). Also, it might be worth adding to the
cupyx name space given that the Autograd project is inactive now?
[Emilio, chainer] autograd & cupy are orthogonal to each other,
Mostly you can think of chainer as an autograd library that has components for nns already included.
As we decided to relegate chainer to maintenance mode, we won't be adding or maintaining any kind of autograd support for cupy.