Where communities thrive

Join over 1.5M+ people
Join over 100K+ communities
Free without limits
Create your own community

Explore more communities

AccelerateHS/Lobby

http://www.acceleratehs.org

People

Repo info

Activity

cheater

@cheater

oh :(

Tom Smeding

@tomsmeding

it seems the nf makes a massive difference...

aaaaand it gets killed by the oom killer

cheater

@cheater

yeah nf makes a massive difference

the results look more normal

@tomsmeding try lowering dbPartNum

regarding the oom

say from 500k to 200k

or 100k

Tom Smeding

@tomsmeding

my gtx 1050 mobile is 2.1x as fast as my i7-700HQ for dbPartNum = 10000

seems to work :)

tomsmeding @tomsmeding will be making dinner

cheater

@cheater

yeah :)

Jonathan Fraser

@JonathanFraser

Any Ideas? Try to get this spun up on the new GPU support for WSL2.

Jonathan Fraser

@JonathanFraser

Realized I had the 11.3 toolkit installed, downgraded to 10.2 but now there's linking issues

Jonathan Fraser

@JonathanFraser

Tried setting ld_library_path but no joy. Anyway to see the full list of ldflags in use?

Jonathan Fraser

@JonathanFraser

So it looks like the problem is it's only using the cuda tookit directory as a library location and that doesn't contain a libcuda.so by default (there is a stub impl in ./stub/libcuda.so)

if I symlink the system libcuda into the toolkit things seem to work

Trevor L. McDonell

@tmcdonell

hey @JonathanFraser, long time no see! how are things?

I don’t have much experience on windows but glad to see you seem to have it working. There is a PR for CUDA 11 support still waiting for me to merge and release. Not sure if that would have avoided the problem for you, I’ll try and get around to that soon anyway

Trevor L. McDonell

@tmcdonell

@JonathanFraser btw I uploaded a new version of the cuda bindings which works with cuda-11

Callan McGill

@Boarders

Completely random question (not a feature request, just curious!): how feasible would it be to use something like vulkan for the accelerate backend? Is that language just way too far away accelerate at present / cuda offering specific features or is it just a case of "wow that is a lot of elbow grease"?

Trevor L. McDonell

@tmcdonell

just a lot of elbow grease

vulkan uses llvm internally (just like cuda), so what we'd really want to do is munge the llvm we generate so that the vulkan driver accepts it, and just use the vulkan api as a coordination layer (launching kernels, allocating device memory, etc.)

(saying "munge the llvm we generate" seriously understates the work required, but still...)

there are plans to support more devices/backends, but it's a lot of elbow grease ):

Callan McGill

@Boarders

ok cool! Nice to know (if only for my curiosity). Accelerate is such a cool project and it is amazing what you have accomplished imo

Trevor L. McDonell

@tmcdonell

thanks! it's always super nice to hear from users ^_^

Troels Henriksen

@athas

Vulkan compute was seriously complicated last I looked at it. Conceptually doable, but the boilerplate is apocalyptic.

Vulkan is mainly for graphics, and heavily stratified so that it is implementable on very basic devices that cannot do general-purpose computation. You have to enable an enormous number of "extensions" to get access to thing like a standard notion of memory.

One thing we ran into is that Vulkan did not support "bitcasts" (without extensions that were not universally supported), which is necessary if you want to reuse memory for different types.

But this was two years ago, and it was pretty clear that things would get better... eventually.

Trevor L. McDonell

@tmcdonell

ah interesting! maybe it's better these days, but yeah I'm not sure it would be the best target to go after (not that I have the time...)

Johannes Maier

@kenranunderscore

Hi everyone! I'm trying to get the accelerate-examples to run, preferable with the docker variant that is to be found in the top-level directory. To successfully perform a stack build I had to recreate the symlink to libcuda.so.1 that had been removed again during the Docker stages. Am I doing something wrong that I even need libcuda inside the container?

Johannes Maier

@kenranunderscore

Oh, and when I try to run one of the examples inside the container, I'm seeing

/opt/accelerate-examples/.stack-work/install/x86_64-linux/15ab62d7d6829823d56e91b6a792b2fece77c32bb61dfca3848225d46a7ad9b4/8.10.2/bin/accelerate-canny: error while loading shared libraries: libffi.so.7: cannot open shared object file: No such file or directory

, which makes me suspect even more that I'm simply using the Docker setup the wrong way. Any help is greatly appreciated!

Johannes Maier

@kenranunderscore

In the meantime I've read the README in accelerate-llvm, where I learned that I should probably use NVIDIA-Docker. Now I've tried to get the examples to run locally (I have the CUDA toolkit installed on Ubuntu, as well as the other dependencies) but I'm still getting errors. Will try again sometime later this week with more questions :)

Trevor L. McDonell

@tmcdonell

Hi @kenranunderscore. Yes, to use the docker images you’ll probably want to run via nvidia-docker, which should take care of the libcuda.so.1 error.

Feel free to post any errors you run into, I’ll do my best to respond although might not be too quick as I’m on leave right now. You might have better luck with the version of accelerate-llvm from GitHub by the way, as that should work with the latest ghc/llvm/CUDA

Charles Durham

@fabricatedmath

@tmcdonell Hey Trevor! Been away from using Accelerate for a while. Back at it and really liking all the changes. Pattern I1, I2 is by itself a huge time savings! However, I'm trying to generate curands on device through FFI and get that device ptr managed by Accelerate. I'm following this stack overflow: https://stackoverflow.com/questions/48526958/how-to-use-a-cuda-deviceptr-as-an-accelerate-array and it seems like the API entirely changed. How should I go about doing this now?

Trevor L. McDonell

@tmcdonell

hey @fabricatedmath, long time no see, how have you been? yes, lots of changes, I’ll try and whip up an equivalent example for you with the new api shortly

Charles Durham

@fabricatedmath

@tmcdonell Good! Keeping busy. Chasing a toddler around also now. How have you been?

After I posted my question I did see a lot of the modules are still importable (I thought they were hidden from import because they were hidden from haddock). Is accelerate-fft the gold standard now for an cuda ffi?

Trevor L. McDonell

@tmcdonell

@fabricatedmath whoa, congratulations! I bet a toddler will indeed keep you busy (: I’m okay, was mostly just busy with teaching before the summer break, taking a bit of time to recover

yep accelerate-fft is a good one to look at, and also accelerate-blas, which has a few more examples

Charles Durham

@fabricatedmath

@tmcdonell Hey, thanks! I seem to have gotten it all working based on those two templates, thanks for the help.

I'm also trying to add a native implementation using MWC randoms and wrapping the call to randomArray into a "Par Native (Future (Vector Float))" but the Array types aren't working together. How do I move the Array from Data.Array.Accelerate into the one from Data.Array.Accelerate.Representation.Array?

Trevor L. McDonell

@tmcdonell

oh I forgot to mention, I have this random number generator which might work for you? https://github.com/tmcdonell/sfc-random-accelerate

the interface is pretty minimal at the moment…

can you share the (non-working) code? might be easier for me to see

Charles Durham

@fabricatedmath

I'll take a look, thanks.

Hey, yeah, here's a gist https://gist.github.com/fabricatedmath/90e64295975288161c0bdbf8f5662692.

The ptx part compiles and runs just fine, the native is giving the error in the gist

Charles Durham

@fabricatedmath

sorry, forgot to @tmcdonell , but please don't mistake this for urgency, really appreciate the help

Chat via Matrix