pesser / iirfilters

IIR filters with thrust.

IIR Filter implementation

Implementation of IIR Filter with thrust. The parallelization enables fully coalesced memory access by transposing the data during the addition of the causal and anti-causal pass with cuBLAS. Furthermore, the implementation can also be used without CUDA enabled devices using the OpenMP backend of thrust. A comparison between different parallelization approaches can be found in the presentation slides (and the different approaches' implementations can be found in the commit history).

Usage

The repository contains two examples, one test and two programs for timing purposes. Calling make builds all of them with the CUDA backend and make omp with the OpenMP backend (make sure to call make clean if you want to __re__build).

Executables

• example_thrust_deriche: Performs Gaussian image blurring using the thrust implementation.

• example_seq_deriche: Same as above with sequential reference implementation.

• test_thrust_deriche: Compares the results of Gaussian image blurring between the thrust implementation and the sequential implementation and gives a warning if the error on any pixel is larger than 1e-4.

• time_thrust_deriche: Performs IIR filter on square data with specified number of rows and prints four columns:

  1. Number of rows
  2. Time for setup (memory allocation and host to device transfer)
  3. Time for horizontal pass
  4. Time for vertical pass
  5. Time for finalizing (device to host transfer)

  Notice that the final version does not use a horizontal pass and columns 3 and 4 should simply be added together and treated as the compute time.

• time_seq_deriche: Same as above with sequential implementation.

Other Files

• coefficients.h: Datastructure and computation of Deriche coefficients used to approximate Gaussian blurring with IIR filter.
• seq_deriche.h: Sequential implementation.
• thrust_deriche.h: Thrust implementation.
• timer.h: Timer that can be used with and without CUDA, makes sure to synchronize with CUDA and returns time in seconds.
• utils.h: Utilities used for debugging and testing.
• collect_sequential.sh: Script to collect timings of sequential implementation for different sizes of input data. Output can be easily read by pandas.
• collect_thrust.sh: Same as above for thrust implementation.
• alternatives/: More alternatives for parallelization to be explored.
• images/: Some example images to be used with the examples.
• png++/: PNG++ - a C++ interface to libpng to read and write png images.
• presentation/: Presentation slides of this work.

Contact

• Judith Massa [e-mail]
• Patrick Esser [e-mail]