cc-metric-collector/collectors/likwid/groups/arm64fx/FLOPS_HP.txt
2021-03-25 14:47:10 +01:00

27 lines
940 B
Plaintext

SHORT Half-Precision MFLOP/s
EVENTSET
PMC0 INST_RETIRED
PMC1 CPU_CYCLES
PMC3 FP_HP_FIXED_OPS_SPEC
PMC4 FP_HP_SCALE_OPS_SPEC
METRICS
Runtime (RDTSC) [s] time
Clock [MHz] 1.E-06*PMC1/time
CPI PMC1/PMC0
HP (FP) [MFLOP/s] 1E-06*(PMC3)/time
HP (FP+SVE128) [MFLOP/s] 1E-06*(((PMC4*128)/128)+PMC3)/time
HP (FP+SVE256) [MFLOP/s] 1E-06*(((PMC4*256)/128)+PMC3)/time
HP (FP+SVE512) [MFLOP/s] 1E-06*(((PMC4*512)/128)+PMC3)/time
LONG
Formulas:
HP (FP) [MFLOP/s] = 1E-06*FP_HP_FIXED_OPS_SPEC/time
HP (FP+SVE128) [MFLOP/s] = 1.0E-06*(FP_HP_FIXED_OPS_SPEC+((FP_HP_SCALE_OPS_SPEC*128)/128))/time
HP (FP+SVE256) [MFLOP/s] = 1.0E-06*(FP_HP_FIXED_OPS_SPEC+((FP_HP_SCALE_OPS_SPEC*256)/128))/time
HP (FP+SVE512) [MFLOP/s] = 1.0E-06*(FP_HP_FIXED_OPS_SPEC+((FP_HP_SCALE_OPS_SPEC*512)/128))/time
-
Half-precision FP rate for scalar and SVE vector operations with different widths. The events for
the SVE metrics assumes that all vector elements are active.