dihydrogen's Introduction

DiHydrogen

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.

License

DiHydrogen is distributed under the terms of the Apache License (Version 2.0).

All new contributions must be made under the Apache-2.0 licenses.

See LICENSE, COPYRIGHT, and NOTICE for details.

SPDX-License-Identifier: Apache-2.0

LLNL-CODE-800100

dihydrogen's People

Contributors

Stargazers

Watchers

dihydrogen's Issues

[H2Core] Logger flags require MPI linkage

The logic for using MPI in the Logger flags is actually incorrect. Finding the header is not sufficient as MPI linkage is not typically included "for free" with most compilers on most systems.

We can actually just tie right in to the DistConv MPI logic for now. Eventually H2 will require MPI, at least indirectly.

[DistConv] MIOpen path through convolution algorithm selection

Currently, any value for the "name" passed to get_{fwd,bwd_{data,filter}}_algorithm() (e.g., here) that is not "AUTOTUNE" will probably fail for one reason or another.

"HEURISTIC" will fail because this isn't supported in MIOpen. Other options will fail because MIOpen requires that their Find APIs be called, and we don't currently incorporate that when the algorithm is not "AUTOTUNE".

[DistConv] Debug builds fail with ROCm enabled

Looks like some issues with assert. The following TUs are failing:

legacy/tests/tensor/CMakeFiles/test_halo_exchange_cuda.dir/test_halo_exchange_cuda.cu.o
legacy/tests/tensor/CMakeFiles/test_tensor_mpi_cuda_shuffle.dir/test_tensor_mpi_cuda_shuffle.cu.o
legacy/tests/tensor/CMakeFiles/test_tensor_mpi_cuda_copy.dir/test_tensor_mpi_cuda_copy.cu.o

It looks like there are some assert() calls in __host__ __device__ functions and the HIP compiler is unhappy about that.