Releases: LLNL/DiHydrogen
Releases · LLNL/DiHydrogen
v0.3.0
========================= 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
v0.2.0
v0.1
============================== 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