SPECaccel 2023: Changes from SPEC ACCEL V1.4

(To check for possible updates to this document, please see http://www.spec.org/accel2023/Docs/ )

Introduction: Who Wants SPECaccel 2023

SPECaccel 2023 (aka version 2.0) is an moderate update to SPEC ACCEL V1.4. While a still widely used suite, especially to test compiler's directive based offload model support, the runtimes on modern accelerators became too short to measure performance. The primary motivation of v2.0 is to update the benchmark workloads to use more memory and run longer to allow for better comparisions.

Overview of changes in v2.0:

Change the name of the benchmark suite from "SPEC ACCEL" to "SPECaccel" to better align with naming of other SPEC Benchmarks.
Driver renamed from "runspec" to "runaccel"
Add new larger and longer running ref workloads. Memory size went from a max of 2GB to 15GB.
Add "metadirective" to OpenMP to select one of three schedules. "src.alt" versions of each are available for compilers that do not support metadirective.
Add an OpenMP "loop" schedule to each benchmark, which is also the default.
Drop PALM (351/551) and CG (354/554). With the larger workload, over 50% of the runtime was spent on the host with these benchmarks. Given SPECaccel is focused on Accelerator performance, these benchmarks no longer represented the goals of the suites.
Drop MRI-Q (314/514). Workload was based on a MRI input and no suitable larger input set could be found.
Drop OpenCL suite. The fixed workloads were difficult to develop and there was not a strong motivation to update the suite.
Merged OpenACC and OpenMP into a single source base with users being able to select the model they wish to use via the new "pmodel" option. The merged source also ensures direct compariblity as the sources are identical sans the directives.

Note that other than the addition of metadirectives, small fixes, merging OpenACC and OpenMP, and an OpenMP "loop" schedule, the source remains largely unchanged from v1.4