About

A lightweight, standalone Makefile-based wrapper for SPEC CPU 2026 rate benchmarks. It copies sources and data from a SPEC CPU 2026 installation and builds/runs them independently using plain Makefiles, without requiring the SPEC harness at runtime (except for the specdiff validation tool).

Includes all 26 specrate benchmarks:

SPECint 2026 rate (14)	SPECfp 2026 rate (12)
706.stockfish_r	709.cactus_r
707.ntest_r	722.palm_r
708.sqlite_r	731.astcenc_r
710.omnetpp_r	736.ocio_r
714.cpython_r	737.gmsh_r
721.gcc_r	748.flightdm_r
723.llvm_r	749.fotonik3d_r
727.cppcheck_r	765.roms_r
729.abc_r	766.femflow_r
734.vpr_r	767.nest_r
735.gem5_r	772.marian_r
750.sealcrypto_r	782.lbm_r
753.ns3_r
777.zstd_r

How to

• Set SPEC CPU 2026 path in env vars

export SPEC=/path/to/CPU2026

• Source shrc

cd $SPEC && source shrc

• Copy source

cd $SPEC_LITE
make copy_allr

• Copy data

make copy_data_all -j `nproc`

• Compile binaries

export ARCH=riscv64
export CROSS_COMPILE=riscv64-unknown-linux-gnu-
make build_allr -j `nproc`

• Collect result (optional)

bash scripts/collect.sh

• Run on localhost (optional)

make run-all-refrate # Don't use -j `nproc` here for single thread benchmarks
make report-int-refrate
make report-fp-refrate

You can also specify VALIDATE=0 and REPORT=0 to disable validation and report generation to speed up the running process, this relaxes the need of SPEC env var at runtime:

make ARCH=riscv64 VALIDATE=0 REPORT=0 run-all-test

When running on real hardware, sometimes you may want to know the performance counters of the binaries, you can specify PROFILER argument to run the compiled binaries with a profiler such as perf:

make ARCH=riscv64 run-int-test PROFILER="perf stat -e cycles,instructions,branch-misses,cache-misses --append -o ../../perf.log"

When you need to compile different binaries with different architecture or flags, you can specify TAG argument to distinguish the compiled binaries, it will use build$(TAG) as the build folder name:

make ARCH=x86_64 build_allr -j `nproc`
make ARCH=x86_64 run-int-test
make ARCH=riscv64 TAG=-riscv64 build_allr -j `nproc`
make ARCH=riscv64 TAG=-riscv64 run-int-test

When you need to run multiple compiled binaries on shared storage (e.g. NFS) at the same time, you can specify RUN_TAG argument to distinguish the run folders, it will use run$(RUN_TAG) as the run folder name:

make ARCH=riscv64 RUN_TAG=-run1 run-int-test
make ARCH=riscv64 RUN_TAG=-run2 run-int-test

Note for Fortran benchmarks

Some Fortran benchmarks (e.g. 765.roms_r) may require ulimit -s unlimited before running to avoid stack overflow, especially when compiled with LLVM's flang-new.

Note for astcenc

The 731.astcenc_r benchmark has a uninitialized variable on the stack which can cause non-deterministic behavior and runtime SIGSEGV on some platforms with some compilers (e.g. GCC-16 on x86-64). You can use patch -p1 < patches/optional/731-astcenc-init-partitions.patch to avoid this issue.

Note for gmsh

The 737.gmsh_r benchmark bundles Shewchuk's robust geometric predicates (predicates.c), which depend on the exact two-step rounding of a*b+c. Whenever the target exposes FMA (e.g. -march=x86-64-v3, -march=x86-64-v4), GCC-16 contracts those expressions into a single fused-multiply-add and breaks the round-off tracking — typical symptoms are a segmentation fault on data/refrate/input/gasdis.geo and a non-terminating run on Torus.geo. The Makefile compiles only predicates.c with -ffp-contract=off to preserve the algorithm's preconditions; perf impact on the rest of the benchmark is nil (measured < 0.5% on the train workload on Zen3, Zen5 and Raptor Lake). You can use patch -p1 < patches/optional/737-gmsh-disable-opt.patch to enable this.

Note for femflow

The 766.femflow_r benchmark has many large inline functions, but compiler may not inline them due to the cost model. We recommand adding -DSPEC_INLINE_POLICY=SPEC_INLINE_LOOSE to the CFLAGS when compiling femflow to make inline functions to be inlined more aggressively, which can significantly improve the performance in some cost models (e.g. GCC-16 with -O3 on x86-64). You can use patch -p1 < patches/optional/766-femflow-inline.patch to enable this.

Note for lbm on AMD Zen 5 with GCC 16

We discovered 782.lbm_r may have 2x performance difference on AMD Zen 5 depending on PC alignment. When comparing the performance of AMD Zen 5 with other CPUs using GCC-16 with -O3 -march=x86-64-v4. It's recommand to use patch -p1 <patches/optional/782-lbm-zen5-prefetch-conflict.patch to avoid the prefetch conflict issue which can cause significant performance degradation. For other compilers or different flags on Zen 5, the align issue may be different and need further manual tuning.

Reference

• https://github.com/OpenXiangShan/CPU2006LiteWrapper
• https://github.com/OpenXiangShan/CPU2017LiteWrapper