Skip to content

Latest commit

 

History

History
118 lines (78 loc) · 3.87 KB

README.md

File metadata and controls

118 lines (78 loc) · 3.87 KB
Raw

Tensor Transpose Compiler

The Tensor Transpose Compiler (TTC) generates high-performance parallel and vectorized C++ code for multidimensional tensor transpositions.

TTC supports arbitrarily dimensional, out-of-place tensor transpositions of the general form:

ttc

where A and B respectively denote the input and output tensor; represents the user-specified transposition, and and being scalars (i.e., setting != 0 enables the user to update the output tensor B).

Please also have a look at TTC-C which provides a C wrapper API for TTC.

Current version: v0.1.1

Attention: Please note that TTC is no longer updated since it has been replaced by an equivalent C++ library: HPTT.

Key Features

  • Generates parallelized and vectorized code
  • Support for multiple instruction sets: AVX, AVX2, AVX512, KNC, CUDA
  • Support for all datatypes (i.e., single, double, single-complex and double-complex)
  • Support for mixed precision (i.e., different data types for A and B)
    • For instance, this feature can be used to generate a mixed-precision BLAS
  • Support for multiple leading dimensions in both A and B
    • This enables the user to extract (and transpose) a smaller subtensor out of a larger tensor
  • TTC allows the user to guide the code generation process:
    • E.g., specifying the --maxImplementations=N argument will limited the number of generated implementations to N

Install

  1. Clone the repository into a desired directory and change to that location:

    git clone https://github.com/HPAC/TTC.git cd TTC

  2. Install TTC:

    python setup.py install --user

  3. Make sure that you export the TTC_ROOT environment variable (add to your .bashrc):

    export TTC_ROOT=pwd

  4. You might have to add the installed location to your PATH environment variable:

    export PATH=$PATH:~/.local/bin

Getting Started

Please run ttc --help to get an overview of TTC's parameters.

Here is one exemplary input to TTC:

ttc --perm=1,0,2 --size=1000,768,16 --dataType=s --alpha=1.0 --beta=1.0 --numThreads=20

Requirements

You must have a working C compiler. I have tested TTC with:

  • Python (tested with v2.7.5 and v2.7.9)
  • GCC (>= 4.8)
  • Intel's ICC (>= 15.0)

Benchmark

TTC provides a benchmark for tensor transpositions.

python benchmark.py <num_threads>

This will generate the input strings for TTC for each of the test-cases within the benchmark. The benchmark uses a default tensor size of 200 MiB (see _sizeMB variable)

Citation

In case you want refer to TTC as part of a research paper, please cite the following article (pdf):

@article{ttc2016a,
   author      = {Paul Springer and Jeff R. Hammond and Paolo Bientinesi},
   title       = {{TTC}: A high-performance Compiler for Tensor Transpositions},
   archivePrefix = "arXiv",
   eprint = {1603.02297},
   primaryClass = "quant-ph",
   journal     = {CoRR},
   year        = {2016},
   issue_date  = {March 2016},
   url         = {http://arxiv.org/abs/1603.02297}
}

Changelog

V0.1.1:

  • Improved performance for streaming stores (only applicable if beta=0)

Feedback & Contributions

We are happy for any feedback or feature requests. Please contact springer@aices.rwth-aachen.de.

We also welcome any contributions to the code base.