BasisSampler

Installation

julia> Pkg.clone("git@github.com:samuelpowell/BasisSampler.jl.git")

julia> Pkg.build("BasisSampler")

Usage

Express some arbitrary function with an (optionally sparse) vector of basis coefficients:

julia> v = rand(1000)*2-1

Build a Sampler:

julia> s = Sampler(v)

Draw unbiased samples representing the index of the basis coefficient, and the polarity of that sample:

julia> sample(s)
(846, -1.0)

Reconstruct a function:

julia> function recon(n)
  l = length(v)
  r = zeros(l)

  @inbounds for i in 1:n
    e, w = sample(s)
    r[e] += w
  end

  r .*= absnorm(s)/n

  println("Norm: ", norm( r - v )./l)

  end
recon (generic function with 1 method)

julia> recon(1e6)
Norm: 0.0005198154172866841

CUDA Module (experimental)

BasisSampler.jl provides a PTX module and helper types which can be used by a CUDA application. The PTX module is compiled automatically during the package build process if CUDAdrv.jl is installed, and nvcc is available on the path.

To call the kernel function, include src/CUDA/sampler.h in your project, link against src/CUDA/sampler.ptx using the CUDA driver JIT features, and pass the appropriate structures to your CUDA application.

To upoad the neccedary data to the GPU:

julia>  ds = CuSampler(s)

Then pass a CuSamplerStruct type to your kernel, acquired by calling custruct on the previously created object :

julia> cudacall(..., (...,{CuSamplerStruct},...), ..., custruct(ds), ...)

Note that the structure contains only an integer and a set of pointers to the GPU, so this can be safely passed by value.