Skip to content

Peakflops armv8

Thomas Gruber requested to merge peakflops_armv8 into master

Created by: JanLJL

Created peakflops benchmarks for likwid-bench on ARMv8. This includes

  • Benchmarks for SP and DP operands
  • A combination of MUL and ADD, or pure FMA instructions
  • Scalar, SVE128, SVE256, and SVE512 versions

When running those on a single core as likwid-bench -t BENCHMARK -w S0:40kB:1 and comparing the result with the expected peak performance, all benchmarks reach more than 99% of the theoretical peak performance:

benchmark               measured_MFLOPS expected_MFLOPS relative_performance_%
peakflops               3587.83         3600.0          99.66
peakflops_fma           7174.63         7200.0          99.65
peakflops_sve128        7170.39         7200.0          99.59
peakflops_sve128_fma    14340.99        14400.0         99.59
peakflops_sve256        14315.83        14400.0         99.42
peakflops_sve256_fma    28631.60        28800.0         99.42
peakflops_sve512        28527.05        28800.0         99.05
peakflops_sve512_fma    57035.30        57600.0         99.02
peakflops_sp            3589.91         3600.0          99.72
peakflops_sp_fma        7180.22         7200.0          99.73
peakflops_sp_sve128     14341.46        14400.0         99.59
peakflops_sp_sve128_fma 28628.94        28800.0         99.41
peakflops_sp_sve256     28633.55        28800.0         99.42
peakflops_sp_sve256_fma 57045.71        57600.0         99.04
peakflops_sp_sve512     57051.17        57600.0         99.05
peakflops_sp_sve512_fma 114103.75       115200.0        99.05

Merge request reports