MAGMA Downloads

MAGMA provides implementations for CUDA, HIP, Intel Xeon Phi, and OpenCL. The latest releases are MAGMA 2.5.3, hipMAGMA 2.0.0, MAGMA MIC 1.4.0, and clMAGMA 1.3, respectively. The libraries available for download are listed below in the order of their release dates.

Please use any of the following publications to reference MAGMA.

MAGMA Bitbucket repository:


Showing 1-5 of 55 Entries
MagmaDNN 1.2

MagmaDNN 1.2 is now available. MagnaDNN provides HP data analytics and machine learning tools using MAGMA as its computational backend. Updates in this release include:

  • Bug fixes and performance improvements;
  • oneDNN (MKL DNN) support including fully connected, convolutional, and pooling layers;
  • CMake build system;
  • Added NN examples including CNN, ResNet, AlexNet, LeNet, MNIST and CIFAR interactive, and VGG16;
  • Added examples for Tensor Math;
  • C++ style formatter;
  • Modularized distributed optimizer;
  • CIFAR10, CIFAR100, and MNIST data loaders;
  • CUDA streams added;
  • Model summary prinout;
  • Spack package manager installation support.

More information on MagmaDNN 1.2 is given in this paper and tutorial.

Cite as:

Daniel Nichols, Rocco Febbo, Florent Lopez, Kwai Wong, Stanimire Tomov, and Jack Dongarra. (2020, July 30). MagmaDNN (Version 1.2). Zenodo.

MagmaDNN's repository is on Bitbucket:

release-magmadnn-v1.2.tar.gz   Download View License

hipMAGMA 2.0.0

hipMAGMA 2.0.0 is now released. This is the second release of hipMAGMA, providing support to AMD GPUs. Updates include:

  • Merged and generated from the latest MAGMA 2.5.3 release;
  • Provided currently are around 250 LAPACK functions, amounting to about 1,000 routines when counting all precisions. These routines depend on BLAS and are ported almost without change;
  • Provided are also BLAS, batched BLAS, and auxiliary routines that were orinally written in CUDA and ported to HIP;
  • Performance improvements through specific tuning for AMD GPUs in existing BLAS and bug fixes;
  • Improved documentation and installation;
  • Enabled use of the new hipBLAS routines that were added in the ROCm 3.3 and 3.5 releases;
  • Added and optimized TRMM, TRMV, GEMV, SYMV, GEAM, and batched BLAS in all precisions;
  • hipMAGMA 2.0.0 requires ROCm 3.5 or older;
  • More information is available in the hipMAGMA branch, now a Git repository hosted on bitbucket
hipMAGMAv2.0.0.tar.gz   Download View License

MAGMA 2.5.3

MAGMA 2.5.3 is now released. Updates include:

  • Enhancements to enable hipMAGMA generation from MAGMA to support AMD GPUs through single (MAGMA sources);
  • New routine: add syrk in all precisions (needed for hipMAGMA);
  • New routine: add hemm/symm in all precisions (needed for hipMAGMA);
  • New routine: add GEMM-based herk and her2k in all precision (needed for hipMAGMA);
  • Bug fix in cmake when USE_FORTRAN is OFF;
  • Bug fix in example_sparse.c;
  • Fix support for half computation in magmablas_hgemm_batched tester for CUDA < 9.2.
magma-2.5.3.tar.gz   Download View License

hipMAGMA 1.0.0

hipMAGMA 1.0.0 is now released. This is an initial release for the HIP runtime to support AMD GPUs.

  • hipMAGMA is based on MAGMA 2.5.2 to provide the MAGMA 2.5.2 functionalities to AMD GPUs. All CUDA-based sources in MAGMA 2.5.2 are converted to HIP. Tar-ball is available below or use direct download from the hibMAGMA branch
    hg clone ssh:// -r hipMAGMA
  • Scrips are provided to do the conversion through
    make -f make.gen.hipMAGMA
  • Specific tuning for AMD GPUs is added for some BLAS routines
  • Added SYMM, SYRK, and SYR2K for all precisions (previously not available in MAGMA and not yet in rocBLAS) as needed to enable solvers for generalized symmetric/Hermitian-definite eigenproblems
  • This is an initial release and not all BLAS functionalities are available
  • More information is available in the hipMAGMA branch
hipMAGMAv1.0.0.tar.gz   Download View License

MAGMA 2.5.2

MAGMA 2.5.2 is now released. Updates include:

  • New routine: magmablas_hgemm_batched for fixed size batched matrix multiplication in FP16 using the Tensor Cores.
    The routine does not currently support pre-Volta GPUs.
    The routine outperforms cuBLAS for sizes less than 100, as well as for general sizes that are not multiple of 8.
    The kernel is tuned for the notrans-notrans case only.
    Comprehensive tuning is planned in future releases;
  • Fix magmablas_?gemm_vbatched routines to correctly handle batch sizes over 65535. The same fix is applied to vbatched syrk, herk, syr2k, her2k, symm, hemm, and trmm;
  • Fix a bug in the FP32 <-> FP16 conversion routines (magmablas_hlag2s and magmablas_slag2h). The bug used to cause a launch failure for very large matrices;
  • Fix a bug in batched LU factorization to avoind NaNs when singularity is ancountered;
  • Fix a bug in batched LU factorization to ensure that the first pivot is always returned even when multilpe pivots with the same absolute value are found;
  • Add Frobenius norm for general matrices
    (supported as option to magmablas_Xlange for X = 's', 'd', 'c', or 'z').
magma-2.5.2.tar.gz   Download View License

Showing 1-5 of 55 Entries

Copyright © 2020 The University of Tennessee. All rights reserved.

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
· Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
· Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer listed in this license in the documentation and/or other materials provided with the distribution.
· Neither the name of the copyright holders nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

This software is provided by the copyright holders and contributors "as is" and any express or implied warranties, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose are disclaimed. in no event shall the copyright owner or contributors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software, even if advised of the possibility of such damage.

Sep 19 2020 Admin Login