Tips, tricks, findings on the NEC SX-Aurora vector engine, programming, administration, system software, projects, news. https://sx-aurora.github.io Erich Focht Normal ncc, nc++, nfort behavior The NEC Vector Engine (VE) compilers offer a large number of ways to optimize loops, vectorize and parallelize them. The actions of the compilers are being controlled by using compiler options (with a global scope) or specific compiler directives (#pragma _NEC in C/C++),... Fri, 19 Jan 2024 00:00:00 +0000 https://sx-aurora.github.io/posts/nec-compilers-intrinsics/ https://sx-aurora.github.io/posts/nec-compilers-intrinsics/ Llama2 with bfloat16 on the Vector Engine Erich Focht This post describes work that builds on and extends the initial port of llama2.c to the VE as described in Llama2 on the SX-Aurora Vector Engine. In that article I showed that two quite simple code changes enable the SX-Aurora VE... Mon, 09 Oct 2023 00:00:00 +0000 https://sx-aurora.github.io/posts/Llama2-on-VE-bf16/ https://sx-aurora.github.io/posts/Llama2-on-VE-bf16/ Llama2 on the Vector Engine Erich Focht The SX-Aurora Vector Engine (VE) is an accelerator for HPC applications with large memory bandwidth and vector registers with the length of 16kbits. It is a long vector system that has been conceived for HPC and supports only vector computation in 32 and... Wed, 23 Aug 2023 00:00:00 +0000 https://sx-aurora.github.io/posts/Llama2-on-VE/ https://sx-aurora.github.io/posts/Llama2-on-VE/ Erich Focht NEC has recently changed its policy for providing the proprietary NEC compiler collection (ncc, nc++, nfort). The need for a compiler license has been eliminated. Compiler support packages are still available and can be purchased through your local NEC sales representative. These compilers are auto-vectorizing and auto-parallelizing and... Sat, 20 May 2023 00:00:00 +0000 https://sx-aurora.github.io/posts/nec-compilers-no-license/ https://sx-aurora.github.io/posts/nec-compilers-no-license/ Erich Focht Proprietary LLVM-IR Vectorizer NEC has released its proprietary vectorizer as a black box binary pluggable into LLVM. With some limitations, though, but following almost all compiler options and compiler directoves described in the ncc/nc++ compier manual. The vectorizer is called llvm-vec and is supposed to be called on... Sun, 15 Jan 2023 00:00:00 +0000 https://sx-aurora.github.io/posts/llvm-vec-docker-image/ https://sx-aurora.github.io/posts/llvm-vec-docker-image/ Presentation at the Workshop for Sustained Simulation Performance, March 16-19, 2021. References and links: Wikipedia Tue, 01 Jun 2021 00:00:00 +0000 https://sx-aurora.github.io/posts/Molecular-Docking-Simulation/ https://sx-aurora.github.io/posts/Molecular-Docking-Simulation/ Erich Focht Heterogeneous programs using vector engine offloading frameworks like VEO or VEDA are running on two architectures: the main program and threads spawned by it run on the host (VH), offloaded “kernels” run on the vector engine(s) (VE). Host Side Debugging Debugging the host side of the program (main)... Tue, 12 Jan 2021 00:00:00 +0000 https://sx-aurora.github.io/posts/aveo-debugging/ https://sx-aurora.github.io/posts/aveo-debugging/ Erich Focht The monitoring tool veperf was introduced in November 2018 in the blog post VE monitoring with veperf. It is part of the py-veosinfo package and displays live performance metrics of own running VE programs. veperf has been extended to display some additional metrics: load and store memory bandwidths,... Sun, 19 Jul 2020 00:00:00 +0000 https://sx-aurora.github.io/posts/veperf-bw-power/ https://sx-aurora.github.io/posts/veperf-bw-power/ Simon Moll Introduction LLVM-VE-RV 1.5.1 is a new development release of the LLVM-VE compiler for SX-Aurora. This release offers the following key features: Vector code generation and the ability to leverage LLVM’s vectorizers for automatic vectorization. Early support for VH-to-VE OpenMP target offloading for C programs, based on sotoc by... Thu, 09 Apr 2020 00:00:00 +0000 https://sx-aurora.github.io/posts/llvm-ve-rv/ https://sx-aurora.github.io/posts/llvm-ve-rv/ Erich Focht Introduction In the past weeks I re-implemented the VEO API by using a completely different communication method between vector host (VH) and vector engine (VE) than the original veoffload project. Instead of the system call semantics for controlling VE processes AVEO uses a little RPC layer built on... Wed, 18 Mar 2020 00:00:00 +0000 https://sx-aurora.github.io/posts/AVEO/ https://sx-aurora.github.io/posts/AVEO/ Erich Focht The Issue With the update of the ncc/nc++/nfort compilers to the versions 3.0.1 and higher we saw sometimes that a VEO program failed in the veo_load_library() call, reporting that the VE .so library could not be found, for example like this: failed veo_load_library: No such file or directory... Thu, 06 Feb 2020 00:00:00 +0000 https://sx-aurora.github.io/posts/relinking-veorun/ https://sx-aurora.github.io/posts/relinking-veorun/ Erich Focht The little veo-sync library implements tools for synchronizing contexts in VEO. They need the VH functions patched into VEO, i.e. the capability to enqueue not only VE kernels but also VH function calls into a context. This feature is included in VEO versions 2.2.2 and higher. Building and... Wed, 05 Feb 2020 00:00:00 +0000 https://sx-aurora.github.io/posts/veo-sync/ https://sx-aurora.github.io/posts/veo-sync/ Erich Focht This post describes some bits of the tuning activity which lifted the VE HPCG performance from ~67 GFLOPS to 125.5 128.8 GFLOPS on a VE10B and nearly doubled the performance efficiency to 5.99%. Since the first version of this post some more tunings of the setup and optimization... Fri, 01 Nov 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/hpcg-tuning/ https://sx-aurora.github.io/posts/hpcg-tuning/ Erich Focht The py-veo Python API to vector engine offloading VEO allows the easy use of Notebooks for interactively working on the SX-Aurora vector engine. This post describes how Jupyter notebook can be installed and used and shows an example notebook with code running offloaded on the SX-Aurora VE. Jupyter... Tue, 28 May 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/py-veo-jupyter/ https://sx-aurora.github.io/posts/py-veo-jupyter/ Erich Focht Recently the RPM package libsysve has been added a library that can be used to accelerate the file IO bandwidth and latency. The SX-Aurora vector engine (VE) is offloading system calls to its Linux vector host. libveaccio is a quick and easy way with potential to significantly speed... Sun, 31 Mar 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/accelerated-io/ https://sx-aurora.github.io/posts/accelerated-io/ Erich Focht As mentioned in a notice at the beginning of the post Building and Testing LLVM-VE-RV, LLVM has changed the license to Apache 2 while the LLVM-VE repositories at https://github.com/SXAuroraTSUBASAResearch/ were released under the former LLVM license: BSD-2. The change of the license is currently being discussed, until this... Sat, 30 Mar 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/Testing-LLVM-VE-RV-update/ https://sx-aurora.github.io/posts/Testing-LLVM-VE-RV-update/ NOTE: The intrinsics described in this document (usually called ve intrinsics) were deprecated in the LLVM-VE compiler som time during 2020. They have been replaced by variants that include in each vector instruction the vector length for the instruction described by the intrinsic: the vel intrinsics. Carrying the vector length... Wed, 27 Mar 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/ve-llvm-intrinsics/ https://sx-aurora.github.io/posts/ve-llvm-intrinsics/ Keynote at the WPMVP 2019 workshop at PPoPP 2019, Washington, DC, February 16, 2019 The talk introduces the NEC SX-Aurora TSUBASA vector engine in the context of the history of vector computers. References and links: Wikipedia Sun, 17 Feb 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/Vector-Computer-Evolution/ https://sx-aurora.github.io/posts/Vector-Computer-Evolution/ Updated: April 21, 2019 Erich Focht With VEOS 2.0.1 published mid January 2019 we finally get glibc support for VE programs. The glibc based environment should bring some benefits, it is more compatible to Linux and has better and more configurable memory management. We wanted to have glibc right from... Thu, 24 Jan 2019 00:00:00 +0000 https://sx-aurora.github.io/posts/VEOS-2.0.1-update/ https://sx-aurora.github.io/posts/VEOS-2.0.1-update/ Erich Focht Introduction py-veo is a Python module implementing the VEO API for Vector Engine Offloading, AKA veoffload. It allows the use of the VE from Python programs, even interactively and via IPython or Jupyter notebooks. VE “kernels” must be written in C, C++ or Fortran. I wrote py-veo a... Wed, 19 Dec 2018 00:00:00 +0000 https://sx-aurora.github.io/posts/py-veo-by-examples/ https://sx-aurora.github.io/posts/py-veo-by-examples/