Hello. I have been hacking away at this for a few weeks now and still can’t figure out what is going on. I have a non-allocating 1D distance transform (Felzenswalb algorithm) that is easy to parallelize, as you can see below: function transform!(f, output, v, z) z[1] = -Inf z[2] = Inf …

Never mind, this is the fix and this works! function transform!(img::AbstractGPUMatrix, output, v, z; maxworkgroup = 1024) backend = get_backend(img) kernel = transform_kernel_cols(backend) kernel(img, output, v, z, ndrange = size(img, 1), workgroupsize = min(maxworkgroup, size(img, 1))…

Convert a non-allocating distance transform algorithm to GPU using KernelAbstractions.jl

Specific Domains GPU

vchuravy November 7, 2023, 7:13pm 7

You could also not provide a workgroupsize and just let KernelAbstractions choose one for you.

Also note that KA has a copyto! function

Topic		Replies	Views
GPU kernel that is ~20x slower than corresponding CPU version GPU	9	786	November 10, 2023
Thinking through distance matrix calculation GPU gpu , multithreading	5	2788	May 16, 2017
GPU Map without reduction on multiple arrays indices GPU	1	706	February 8, 2019
Using functions in GPU Kernel (via KernelAbstractions.jl) (k nearest neighbor kernel) GPU	1	991	January 25, 2021
Calculate FFT on GPU for every row of a 2D array Performance gpu	2	1042	August 14, 2018

Convert a non-allocating distance transform algorithm to GPU using KernelAbstractions.jl

Related topics