NVlabs/timeloop

Timeloop performs modeling, mapping and code-generation for tensor algebra workloads on various accelerator architectures.

What it solves

Timeloop provides a way to model, map, and generate code for tensor algebra workloads, specifically for dense and sparse tensors, across various accelerator architectures. It helps designers avoid the manual effort of performance and energy projection for different hardware designs.

How it works

The system is composed of two modular components: an analytical model that emulates architecture designs to project performance and energy, and a mapper that searches for the optimal way to map a tensor-algebra problem onto a specific architecture.

Who it’s for

It is designed for researchers and engineers working on accelerator architecture design and tensor algebra optimization.

Highlights

  • Support for both dense and sparse tensor algebra workloads.
  • Stochastic modeling of compressed-sparse tensor algebra (introduced in version 2.0).
  • Support for spatial skews, flattened mappings, and imperfectly-factorized mappings (introduced in version 3.0).

Related

  • Project
    tensorflow/lingvo

    A modular TensorFlow framework designed for building and scaling neural networks, particularly sequence-to-sequence models and giant language models.

  • Project
    ValeevGroup/tiledarray

    A scalable, block-sparse tensor framework for high-performance tensor arithmetic that allows users to compose complex expressions using math-like C++ syntax.

  • Project
    tensorflow/neural-structured-learning

    A TensorFlow framework for training neural networks using structured signals, such as graphs and adversarial perturbations, to improve accuracy with limited labeled data and increase robustness.

  • Project
    microsoft/onnxruntime

    A cross-platform machine learning accelerator that optimizes inference and training for models from various frameworks like PyTorch and TensorFlow.

  • Project
    LongxingTan/Time-series-prediction

    A TensorFlow-based time series package that provides a unified interface for prediction, classification, and anomaly detection using state-of-the-art deep learning models.