February 2021 - Ecp-aurora-sdk-discuss

New version of mpich on JLSE
by Chan-nui, Christopher 19 Feb '21

19 Feb '21

Hello, We are pleased to announce that an update to mpich, drop 39.1, has been deployed on JLSE. module use /soft/restricted/CNDA/modulefiles module avail mpi/aurora_mpich Important: ATS nodes do not currently use GPU acceleration. Release Notes * Compiled using oneAPI Intel C/C++ compilers (icx and icpx) Details * Added native binary compilation for GPU kernels to improve load times (DG1-specific) * Provided RPMs with GPU support disabled (for Arcticus) * Bug fixes -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

New version of compilers/mkl/advisor on JLSE
by Chan-nui, Christopher 05 Feb '21

05 Feb '21

Hello, We are pleased to announce that an updates to the oneAPI compilers, MKL and Advisor have been deployed on JLSE. They are available as oneapi/2020.12.15.005. **** Important note **** Use dynamic linking when mixing Fortran / MKL for this release (sointel64) Compiler: Improved Fortran offload to GPU (e.g. OvO Fortran improved significantly) Improved 5.0 loop construct support Improved offloading with L0 (e.g. L0 back-end performs better than OCL back-end on USM) Improved hybrid use of OpenMP offloading and DPC++/SYCL with L0 back-end Improved performance on ATS for some workloads (~100x improvement for Nekbone) Enabled -O2 optimization for default device compilation MKL: [BLAS] Improve ATS performance of TRSM and small sizes integer/float/half precision GEMM [LAPACK] Introduce new C and Fortran batch API for LU factorization and solve without pivot, {s,d,c,z}getrfnp_batch_strided and {s,d,c,z}getrsnp_batch_strided [LAPACK] Enable C/Fortran OpenMP offload for {s,d,c,z}getrfnp_batch_strided and {s,d,c,z}getrsnp_batch_strided [FFT] Enable C/Fortran OpenMP offload for the FFTW3 C2C interfaces [Sparse BLAS] Fix memory leaks in the OpenCL OpenMP offload for Sparse GEMV Advisor: Addresses issue with Roofline analysis -- Christopher Chan-Nui christopher.chan-nui(a)intel.com

1 0

Re: [Ecp-aurora-sdk-discuss] Using gdb with Intel GPUs
by Muralidharan, Servesh 02 Feb '21

02 Feb '21

Hi Daniel, We have encountered a few issues with the companion kernel module for oneapi-gdb while running on our Iris nodes. The issue can affect regular users running application on these nodes, so we decided to hold off on the roll out. Could you elaborate on your specific use case and we can see if there is an alternative to support it? The gdb you have loaded should work in cases where you don’t need to debug the kernel within the GPU. We have not really tested TUI mode in either this gdb or oneapi-gdb though, so we might need to investigate this further. -- Servesh From: Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss-bounces(a)lists.jlse.anl.gov> on behalf of Arndt, Daniel via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Date: Tuesday, February 2, 2021 at 1:43 PM To: ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> Subject: [Ecp-aurora-sdk-discuss] Using gdb with Intel GPUs Trying to use gdb (from gdb/9.1-gcc-9.3.0-de2udpz) in TUI mode crashes for me with "(gdb) Aborted (core dumped)". Is there another version that should work fine also in TUI mode? Also, I was surprised to not being able to find oneapi-gdb when loading the oneapi module. Is there some other module that provides that to facilitate debugging inside kernels on Intel GPUs? Best, Daniel

1 0

Using gdb with Intel GPUs
by Arndt, Daniel 02 Feb '21

02 Feb '21

Trying to use gdb (from gdb/9.1-gcc-9.3.0-de2udpz) in TUI mode crashes for me with "(gdb) Aborted (core dumped)". Is there another version that should work fine also in TUI mode? Also, I was surprised to not being able to find oneapi-gdb when loading the oneapi module. Is there some other module that provides that to facilitate debugging inside kernels on Intel GPUs? Best, Daniel

1 0