STOMPSelfJoin

This is a GPU implementation of the STOMP algorithm. STOMP takes a time series as input and computes the matrix profile for a particular window size. You can read more at the Matrix Profile Homepage

For additional features and better performance you should use SCAMP

Environment

This base project requires:

• At least version 9.0 of the CUDA toolkit available here.
• An NVIDIA GPU with CUDA support is also required. You can find a list of CUDA compatible GPUs here
• Currently builds under linux with the Makefile.
• Should compile under windows, but untested.

Usage

• Edit the Makefile
  • Volta is supported by default, but if needed set the value of ARCH to correspond to the compute capability of your GPU.
    • "-gencode=arch=compute_code,code=sm_code" where code corresponds to the compute capability or arch you wish to add.
  • Make sure CUDA_DIRECTORY corresponds to the location where cuda is installed on your system. This is usually /usr/local/cuda-(VERSION)/ on linux
  • Also, by default the kernel parameters are optimized for volta only, if you are building for Pascal or earlier, please tune the variables TILE_HEIGHT_ADJUSTMENT and UNROLL_COUNT in STOMP.cu accordingly
  • Some suggested parameters for different architectures are provided in the comments
• make
• STOMP window_size input_file_path output_matrix_profile_path output_indexes_path (Optional: list of device numbers that you want to run on)
• Example:
• STOMP 1024 SampleInput/randomlist128K.txt profile.txt index.txt 0 2
• By default, if no devices are specified, STOMP will run on all available devices

Matlab hook (deprecated)

• The matlab hook is deprecated. It will be left here and is still mostly functional, but it will no longer be updated.
• If you want to use these kernels in Matlab just run make matlab. This will output STOMP.ptx.
• Just point Matlab to STOMP.ptx and STOMP.cu and you should be able to use any of the kernels through the Matlab parallel computing toolbox.
• StompSelfJoinGPU.m is a matlab script which will compute the matrix profile using the GPU. Assuming step 1 was followed.