Dimension dismatch issue

[yuwenchen95]

I tried to parallel the computation that tries to normalize each column of an array:

using LinearAlgebra, SparseArrays, Random, Test
using BenchmarkTools

function normalize_gpu!(
    σ::CuMatrix{T},
) where {T}
    i = (blockIdx().x-1)*blockDim().x+threadIdx().x

    tmp = zero(T)
    m,n = size(σ)
    if i > n 
        return
    end

    @inbounds for j = 1:m
        tmp += σ[j,i]^2
    end
    # tmp = sum(σ[:,i])

    abs_val = sqrt(tmp)

    @inbounds for j = 1:m
        σ[j,i] /= abs_val
    end
    # @. σ[:,i] /= abs_val

    return nothing
end

g = CuArray(rand(2000,3000))
CUDA.@sync @cuda threads=1024 blocks=cld(size(σ,2),1024) normalize_gpu!(g)

However, it returns the error:

ERROR: LoadError: MethodError: no method matching normalize_gpu!(::CuDeviceMatrix{Float64, 1})
Closest candidates are:
  normalize_gpu!(::CuArray{T, 2}) where T at C:\Users\ddt00\.julia\dev\GPUtest\temp.jl:4
Stacktrace:
  [1] macro expansion
    @ C:\Users\ddt00\.julia\packages\GPUCompiler\BxfIW\src\jlgen.jl:0 [inlined]
  [2] codegen_world_age(ft::Type{typeof(normalize_gpu!)}, tt::Type{Tuple{CuDeviceMatrix{Float64, 1}}})
    @ GPUCompiler C:\Users\ddt00\.julia\packages\GPUCompiler\BxfIW\src\jlgen.jl:179
  [3] cached_compilation(cache::Dict{UInt64, Any}, cfg::GPUCompiler.CompilerConfig{GPUCompiler.PTXCompilerTarget, CUDA.CUDACompilerParams}, ft::Type, tt::Type, compiler::Function, linker::Function)
    @ GPUCompiler C:\Users\ddt00\.julia\packages\GPUCompiler\BxfIW\src\cache.jl:24
  [4] macro expansion
    @ C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\compiler\execution.jl:310 [inlined]
  [5] macro expansion
    @ .\lock.jl:223 [inlined]
  [6] cufunction(f::typeof(normalize_gpu!), tt::Type{Tuple{CuDeviceMatrix{Float64, 1}}}; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ CUDA C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\compiler\execution.jl:306
  [7] cufunction(f::typeof(normalize_gpu!), tt::Type{Tuple{CuDeviceMatrix{Float64, 1}}})
    @ CUDA C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\compiler\execution.jl:303
  [8] macro expansion
    @ C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\compiler\execution.jl:104 [inlined]
  [9] top-level scope
    @ C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\utilities.jl:25
 [10] include(fname::String)
    @ Base.MainInclude .\client.jl:476
 [11] top-level scope
    @ REPL[7]:1
 [12] top-level scope
    @ C:\Users\ddt00\.julia\packages\CUDA\p5OVK\src\initialization.jl:171
in expression starting at C:\Users\ddt00\.julia\dev\GPUtest\temp.jl:31

It's confusing that I am passing an array as the argument but it returns the error pointing to a vector type. Anyone has a clue to fix it?

[maleadt]

It's confusing that I am passing an array as the argument but it returns the error pointing to a vector type.

Vectors are arrays, that's not the source of the issue. The problem is that CuArrays get converted to CuDeviceArrays when passing them to kernels, so your ::CuMatrix type specification is wrong. It's better to be less specific, so if you use ::AbstractMatrix{T} it will apply to the CuDeviceMatrix that your array gets converted to.