Skip to content

Releases: LLNL/DiHydrogen

v0.3.0

08 Nov 00:57
Compare
Choose a tag to compare

========================= Release Notes v0.3.0 =========================

API:

  • H2 added a thin GPU portability layer that supports CUDA and ROCm
    (HIP) programming environments.
  • Added a spdlog-based logging capability.
  • Added initial support for (sequential) tensors.

Performance optimizations:

Internal features:

  • MIOpen is supported as a backend for "DNN compute kernels" in
    DistConv.
  • DistConv is supported on ROCm systems.
  • Added support for Gitlab CI.
  • Support for DaCe convolution kernels in DistConv

Build system:

  • Several updates to support ROCm 5.*.
  • Requires CMake >= 3.21.
  • Requires ROCm 5, with the latest available minor release being
    STRONGLY recommended, when targeting ROCm-based platforms. The
    current packages we need in this stack are:
    -- HIP (host)
    -- hipCUB
    -- ROCm-SMI
    -- MIOpen
    -- ROC-tracer

Bug fixes:

  • Fixed a bug in kernel selection for batchnorm gradients in the case of
    overlap/strides.

Retired features:

v0.2.1

10 Mar 22:25
Compare
Choose a tag to compare

Updated to support Aluminum v0.7 API. Aluminum dependency is only
required for DistConv feature.

v0.2.0

10 Mar 22:25
Compare
Choose a tag to compare

Updated to support the Aluminum v0.5 API

v0.1

30 Sep 01:04
5025fd7
Compare
Choose a tag to compare

============================== Release Notes: v0.1 ==============================

Initial release of the DiHydrogen Distributed Tensor Library.

DiHydrogen is the second version of the Hydrogen fork of the
well-known distributed linear algebra library, Elemental. DiHydrogen
aims to be a basic distributed multilinear algebra interface with a
particular emphasis on the needs of the distributed machine learning
effort, LBANN.

Internal features:

  • Support for the DistConv legacy library
  • Added multi-dimensional tensor class
  • Added support for older CUDA architectures v3.0+
  • Use CUB for memory management
  • Support for point-to-point (P2P) library
  • Support for NVSHMEM halo exchange

DistConv Legacy:

  • batchnorm
  • channel-wise softmax
  • concat
  • convolution
  • deconv
  • leaky-relu
  • pooling
  • relu
  • slice
  • softmax

Build system:

  • Require C++14
  • Added support for cuDNN v8

Bug fixes:

  • Fixed build issues with GCC8