December 2020 - Ecp-aurora-sdk-discuss

New version of DAL/daal4py available in JLSE
by Chan-nui, Christopher 28 Dec '20

28 Dec '20

Hello, We are pleased to announce that an update to DAL and daal4py has been deployed in JLSE. It is available as oneapi/2020.12.15.003. To use daal4py: module add oneapi/2020.12.15.003 source $IDPROOT/bin/activate conda activate $AURORA_BASE_ENV To clone a local environment you can install additional packages in: conda create -p $HOME/myenvs/foo --clone $AURORA_BASE_ENV conda activate $HOME/myenvs/foo This release is binary delivery to support MS111 milestone What is new: * Multi-Node Multi-GPU algorithms: K-Means, Low Order Moments, Covariance, Covariance (Online), PCA * Single GPU algorithms: Logistic Regression (online), KNN Regression, SVM Sparse data support, Heterogeneous data support * XGBoost with Intel GPU support for both single GPU and Multi-Node Multi-GPU * Performance gain: DBSCAN, K-Means, Random Forest algorithms Known problems: 1. Patched Scikit-learn for KMeans algorithm will fail on GEN9 systems with following error: “Sycl error: Native API failed. Native API returns: -37 (CL_INVALID_HOST_PTR) -37 (CL_INVALID_HOST_PTR)” As workaround we suggest to use DG1/ATS system for this case. 1. SVM batch example failures on GEN9 systems examples/sycl/svm_batch.py", line 112, in main\n assert np.allclose(predict_result_gpu, predict_result_classic)\nAssertionError\n' As workaround we suggest to use DG1/ATS system for this case. -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

New version of numba available in JLSE
by Chan-nui, Christopher 22 Dec '20

22 Dec '20

Hello, We are pleased to announce that an update to numba has been deployed in JLSE. This version is available in oneapi/2020.12.15.002. To use: module add oneapi/2020.12.15.002 source $IDPROOT/bin/activate conda activate $AURORA_BASE_ENV To clone a local environment you can install additional packages in: conda create -p $HOME/myenvs/foo --clone $AURORA_BASE_ENV conda activate $HOME/myenvs/foo Numba-dppy: * Offload diagnostics. * Controllable fallback. * Add flags to generate debug symbols. * Implementation of np.linalg.eig, np.ndarray.sum, np.ndarray.max, np.ndarray.min, np.ndarray.mean Dpctl: * Utility functions to transform string to device type and back. * New dpctl.dptensor.numpy_usm_shared module containing USM array. USM array extends NumPy ndarray. * A lot of new examples. Including examples of building Cython extensions with DPC++ compiler that interoperate with dpCtl. Dpnp: * Compatible with DPCTL 0.5 version -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

New Aurora SDK available in JLSE
by Chan-nui, Christopher 18 Dec '20

18 Dec '20

Hello, We are pleased to announce that a new Aurora SDK has been deployed in JLSE. This version 2020.12.15.001 includes multiple improvements and bug fixes as well as updated ATS support. You no longer need to specify oneapi-ats as a module. The oneapi module will load the appropriate drivers for your system. module use /soft/restricted/CNDA/modulefiles module add oneapi For additional details please refer to the latest release notes in the Wiki: https://wiki.jlse.anl.gov/display/inteldga/DGA+SDK+Release+Notes Public Documentation ==================== * ALCF oneAPI overview - https://www.alcf.anl.gov/support-center/aurora/oneapi-overview * Intel GPGPU Software Information - https://dgpu-docs.intel.com/ * Intel oneAPI Beta Documentation - https://software.intel.com/content/www/us/en/develop/tools/oneapi/documenta… Please let us know of any issues with the release or the environment settings. -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

New Aurora SDK coming to JLSE today
by Chan-nui, Christopher 18 Dec '20

18 Dec '20

A new Aurora SDK will be installed today at 4:00 PM CST. Running jobs will not be affected. Default SDK modulefiles will be updated at that time. -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

Re: [Ecp-aurora-sdk-discuss] new Kokkos module
by Zhang, Junchao 02 Dec '20

02 Dec '20

Hi, Patrick, Have you used that to build Kokkos-kernels? It would be nice to also provide a kokkkos-kernels module. Thanks --Junchao Zhang On Nov 30, 2020, at 5:21 PM, Steinbrecher, Patrick via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]>> wrote: Dear all, The Kokkos module on JLSE has been updated using the compiler gold release. Compilation and running is the same as before with beta10. To build your application, just do: module load oneapi module load kokkos Use these updated compiler flags: -std=c++17 -Wno-openmp-mapping -fiopenmp -fopenmp-targets=spir64="-mllvm -vpo-paropt-enable-64bit-opencl-atomics=true" -D__STRICT_ANSI__ And link with: -lkokkoscore -lkokkoscontainers and all .o files given by: /soft/restricted/CNDA/sdk/2020.11.30.1/oneapi/compiler/latest/linux/lib/libomp-*.o If you want to run on ATS you still have to use: module load oneapi-ats module load kokkos/beta10_ats Regards, Patrick -- Ecp-aurora-sdk-discuss mailing list Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov<mailto:[email protected]> https://lists.jlse.anl.gov/mailman/listinfo/ecp-aurora-sdk-discuss

1 0