NVIDIA CUDA GPU Acceleration • matrixprofiler

Overview

By default, matrixprofiler compiles as a portable, high-performance CPU package using RcppThread for multi-threading across Linux, macOS, and Windows. This ensures 100% compliance with CRAN policies without requiring proprietary drivers or external hardware dependencies.

For users with NVIDIA GPUs, matrixprofiler provides native CUDA backends for computationally demanding tasks:

rfcp(..., backend = "cuda"): Relative Frequency Contrast Profile. Accelerated candidate sorting, greedy exclusion zone suppression, and Top-K Plato discovery completely on the GPU with up to 55x speedup over CPU-sorting GPU baselines and over 12x speedup compared to single-threaded CPU execution.
mpx(..., backend = "cuda"): GPU-accelerated Matrix Profile for continuous time series.

Hardware and Software Prerequisites

To build and run matrixprofiler with CUDA acceleration, you need:

NVIDIA GPU with Compute Capability >= 6.0 (Pascal, Volta, Turing, Ampere, Ada Lovelace, Hopper, Blackwell).
NVIDIA CUDA Toolkit (version 11.0 or newer) with nvcc accessible in your system $PATH:
```
nvcc --version
nvidia-smi
```
A standard C++17 compiler (g++ or clang++).

Building matrixprofiler with CUDA

Method 1: Using the Included Build Script (Recommended)

From the cloned repository (or unpacked source package), run the provided helper script inst/cuda/build_cuda_rcpp.sh:

# Clone repository
git clone https://github.com/matrix-profile-foundation/matrixprofiler.git
cd matrixprofiler

# Build CUDA objects and install into your R library:
./inst/cuda/build_cuda_rcpp.sh R CMD INSTALL .

# Or if using devtools:
./inst/cuda/build_cuda_rcpp.sh Rscript -e 'devtools::install()'

Method 2: Manual Environment Configuration

If you prefer to compile manually or integrate with custom build pipelines:

# 1. Compile CUDA objects (.cu -> .o) with position-independent code (-fPIC)
CUDA_HOME=${CUDA_HOME:-/usr/local/cuda}
nvcc -std=c++17 -O3 -Xcompiler -fPIC \
     -c inst/cuda/mpx_gpu_poc.cu -o /tmp/matrixprofiler_mpx_gpu_poc.o
nvcc -std=c++17 -O3 -Xcompiler -fPIC \
     -c inst/cuda/rfcp_gpu_poc.cu -o /tmp/matrixprofiler_rfcp_gpu_poc.o

# 2. Export build flags
export MATRIXPROFILER_ENABLE_CUDA=1
export MATRIXPROFILER_CUDA_OBJECT="/tmp/matrixprofiler_mpx_gpu_poc.o /tmp/matrixprofiler_rfcp_gpu_poc.o"
export MATRIXPROFILER_CUDA_LIBS="-L$CUDA_HOME/lib64 -lcudart"

# 3. Install the package
R CMD INSTALL .

Verifying CUDA Acceleration in R

In your R session, call has_cuda() to verify that the active binary was compiled with CUDA support:

library(matrixprofiler)

has_cuda()
#> [1] TRUE

If has_cuda() returns FALSE, the package is running on the standard CPU backend.

Usage Examples

GPU-Accelerated RFCP (Relative Frequency Contrast Profile)

library(matrixprofiler)

# Generate synthetic series
pos_series <- cumsum(rnorm(5000))
neg_series <- cumsum(rnorm(5000))

# Execute on GPU:
if (has_cuda()) {
  result <- rfcp(
    pos_data = pos_series,
    neg_data = neg_series,
    window_size = 100,
    backend = "cuda"
  )
  print(result)
} else {
  message("CUDA is not enabled on this installation. Running on multi-core CPU...")
  result <- rfcp(pos_series, neg_series, window_size = 100, n_workers = 4)
}

Top-K Plato Discovery with GPU Acceleration

if (has_cuda()) {
  top_k <- rfcp_top_k(
    pos_data = pos_series,
    neg_data = neg_series,
    window_size = 100,
    k = 5,
    backend = "cuda"
  )
  print(top_k)
}