Alea gpu missing output file?

Question

Hello I'm working on ML code with alea gpu. Interesting thing is that when I try to run my code on my laptop everything is working fine except is too slow. When I want to run my code on second PC with more powerful GPU it's crashing on exception:

System.Exception was unhandled
HResult=-2146233088
Message=Error opening outputfile 'C:\Users\USERNAME\AppData\Local\Temp\tmp64C4.tmp': No mapping for the Unicode character exists in the target multi-byte code page
Source location stack:
-> at defining runtime64 (sm61,64bit)
Loading method as kernel:
-> Method: AI_test.NeuralNet.Gpu.Activations.Hyperbolic.[VoidGpuActivationKernel(Single[,], Single[], Single[])]
-> InstanceOpt: AI_test.NeuralNet.Gpu.Activations.Hyperbolic
-> Argument.#0: System.Single[,]
-> Argument.#1: System.Single[]
-> Argument.#2: System.Single[]
Getting or loading method as kernel:
-> Method: AI_test.NeuralNet.Gpu.Activations.Hyperbolic.[VoidGpuActivationKernel(Single[,], Single[], Single[])]
-> InstanceOpt: AI_test.NeuralNet.Gpu.Activations.Hyperbolic
-> Argument.#0: System.Single[,]
-> Argument.#1: System.Single[]
-> Argument.#2: System.Single[]

Source=Alea

Stack trace show me that is crashing on this code:

GpuDevice.Launch(_guessKernelFunc, GuessLaunchParam, Neurons, Input, Output);

But I'm sure that is working just fine on my laptop so do I miss something is it cause by PC settings or something? I have installed Cuda 8.0. I tried before Cuda 9.0 but because I had the same error I tried to replace with older 8.0 version which I use on my laptop but it didn't help.

Second question is Does somebody know how to make compiller to cache my kernel and not generate new GPU code every time I try to load the same static kernel? I tried to look at sample but it's not clear to me.

The compiler automatically caches your kernel. You don't have to do anything. — redb, Nov 08 '17 at 14:26
that's not true I think... when I compare results of my ML on Iris dataset and one layer with 100 neurons CPU net is about 20 times faster so it's more than clear that I have problem with kernel caching. Obviously there is something wrong because no CUDA GPU is slow like this. — Shoxik, Nov 08 '17 at 19:10
I guarantee you the kernel is being cached :) I don't know about the missing file, the performance issue is probably because you don't have enough data to process. In ML 100 neurons per layer is close to nothing. The GPU does not get out of bed for those kinds of numbers ;) You're paying the penalty of moving that data into the GPU. I would suggest the MNIST dataset to start with. — redb, Nov 09 '17 at 20:13
Yup now I can see it. I tried mnist data set and ou yeah GPU with comparison to CPU was on light years faster :) — Shoxik, Nov 22 '17 at 19:37

score 0 · Answer 1 · answered Nov 22 '17 at 19:36

So for those people who will get stopped by this problem solution is easy. My user name folder was in format Name Surname and when I created new user with username without whitespace it started working again. This was very tricky problem but finally this is solution for that.

Alea gpu missing output file?

1 Answers1