GPU

RecurrenceMicrostatesAnalysis.jl supports GPU acceleration for computing RMA distributions. The GPU backend is implemented using KernelAbstractions.jl, which enables portable GPU execution across different hardware backends.

The GPU pipeline is implemented via a GPUCore, which defines a single internal kernel used to compute microstate histograms across supported devices.

Data requirements

The GPU backend supports only Float32 data. Therefore, input datasets must be explicitly converted before being used:

using RecurrenceMicrostatesAnalysis, Distributions
data = StateSpaceSet(Float32.(rand(Uniform(0, 1), 1000)))

1-dimensional StateSpaceSet{Float32} with 1000 points
 0.14377917
 0.67664564
 0.075778745
 0.6407332
 0.66440666
 0.8474236
 0.1413314
 0.53737825
 0.18213958
 0.2772973
 ⋮
 0.33788076
 0.30877522
 0.060387507
 0.06340307
 0.10326237
 0.17285343
 0.5770751
 0.87220806
 0.06731922

Recurrence expressions and metrics

When defining a RecurrenceExpression for GPU execution, the distance metric must be a subtype of GPUMetric. Metrics from Distances.jl are not supported on GPU.

For example:

expr = Standard(0.27f0; metric = GPUEuclidean())

Standard{Float32, GPUEuclidean}(0.27f0, GPUEuclidean())

Moving data to GPU memory

To enable GPU computation, the data must be transferred to GPU memory. For example:

• Using CUDA:

using CUDA
gpu_data = data |> CuVector 

• Using Metal:

using Metal
gpu_data = data |> MtlVector

Output handling

Results computed on the GPU are automatically transferred back to the CPU:

• histogram returns a Counts object.
• distribution returns a Probabilities object.

No manual data transfer is required for the output.

Metrics for GPU

Since the GPU backend does not support Distances.jl, distance metrics must be implemented explicitly as subtypes of GPUMetric.

GPUMetric <: Metric

Implemented GPU metrics

GPUEuclidean <: GPUMetric