Now I want to parallerize something like the following dual coordinate descent method. I’m trying to use CUDA.jl, however I’m stuck for days… # X = (x[1], ..., x[numvec]) numvec = 500000 lenvec = 50000 xs = [sprandn(lenvec, 0.01) for i in 1:numvec] # w is the parameter which I want to optimize. w…

Your assignment to d might happen after some threads have read their value. You need calls to synchronize_thread to prevent that. Do note that’s only possible because you’re launching a single block – when using multiple blocks it is impossible to synchronize threads – which isn’t ideal for performa…

How to parallerize dual coordinet descent mehods on GPU using CUDA.jl?

Specific Domains GPU

msekino May 29, 2020, 10:56pm 7

Thank you so much, it works!
I will try it on the real data X and compare the running time with CPU!

Topic		Replies	Views
CUDA.jl - Atomic Addition Error To A Sparse Array Inside CUDA Kernel GPU cuda , error , memory-allocation , sparse , atomic	4	711	March 31, 2022
Accessing array elements too slow? GPU	10	602	April 23, 2021
CUDA atomic add for ForwardDiff duals? GPU question , gpu , cuda , forwarddiff , atomic	2	974	August 19, 2022
Use of linear operators on a GPU GPU question	3	515	January 18, 2023
Correct utilisation of CUDA kernel for simulations GPU	16	585	February 13, 2025

How to parallerize dual coordinet descent mehods on GPU using CUDA.jl?

Related topics