GPU / CUDA
seany
seany at uni-bonn.de
Wed Mar 17 19:02:58 UTC 2021
Hi
I am looking for a way to parallelize my machine learning model
in D.
I have ~100000 neurones, and just the initialization is very tme
consuming.
I use the pseudocode :
foreach( some condition) {
foreach (other condition) {
allocate a neuron to detect a feature within a segment of an
larger image.
}
}
The allocation is successful. But excruciatingly _slow_ -
emphasis added.
So I tried to get started with dcompute.
My dub.json :
{
"authors": [
"sean"
],
"copyright": "Copyright © 2021, sean",
"dependencies": {
"dlib": "~>0.20.0",
"dcompute": "~>0.1.0"
},
"description": "cuda test",
"license": "proprietary",
"name": "cudatest"
}
My code :
@compute(CompileFor.deviceOnly) module mykernels;
import ldc.dcompute;
@kernel void mykernel(GlobalPointer!float a,GlobalPointer!float
b, float c)
{
*a = *b + c;
}
I run it with : dub build --compiler=ldc2
The result is : /usr/include/d/std/array.d(2975,13): Error:
TypeInfo cannot be used with -betterC (using latest opensuse
version for ldc)
I saw this :
https://forum.dlang.org/thread/amcxssclbfaczrgqjqeo@forum.dlang.org
And this :
https://speakerdeck.com/shigekikarita/grain-d-language-for-deep-learning?slide=17
I would like to have a step by step guid to handle a pair of
foreach loops in dcompute please.
Thnak you
More information about the digitalmars-d-ldc
mailing list