Decimal128 by Miguel

Zig Benchmark Results

IEEE 754-2019 compliant decimal128 high-performance software solution created by Miguel.

Living document — as-measured results. Category codes, profiles, modes & method: Benchmark Key.

This is the Zig view of decimal128 as-measured, band by band, with explicit ratios. It opens with the realistic financial-mix (P-fin) headline, then the per-operation band characterization (P-gen) and FMA. In Zig, d128 is measured against the libbid universal reference (Zig has no in-language decimal peer). It is data only — the categories, magnitude profiles, units, and methodology are defined in the Benchmark Key (and, authoritatively, BenchmarkMatrix.md). The cross-port d128 band-shape matrices (all ports, no alternatives) live in Port-Comparison Benchmark Results; the full index of per-language pages is on the Benchmarks hub.

Summary — Ratio Range by Operation

The ratio for Zig’s reference library on x86_64 (Intel i9-9880H): ratio = libbid / Miguel (> 1× ⇒ d128 faster), broken out by operation. Zig has no in-language decimal peer.

  Add Subtract Multiply Divide
ratio = libbid / Miguel 4× 3× 2.1× – 8× 0.8× – 1.6×

FinMix — realistic financial mix (P-fin)

The headline: one realistic 64-bit financial operation mix — a MIX add/sub stream, mul CP/WP, div CD/WD/ET/PT — ratio = alt / Miguel (> 1 ⇒ d128 faster). This is the profile closest to real financial code.

arm64 (M3 Pro).

port op cat profile arch mode ns alt alt ns ratio run notes
zig add MIX P-fin arm64 thru 2.02 libbid 10.72 5.31× Rzgsw2  
zig sub MIX P-fin arm64 thru 2.50 libbid 11.80 4.72× Rzgsw2  
zig mul CP P-fin arm64 thru 1.43 libbid 23.57 16.48× Rzgsw2  
zig mul WP P-fin arm64 thru 18.64 libbid 34.52 1.85× Rzgsw2  
zig div CD P-fin arm64 thru 30.44 libbid 35.07 1.15× Rzgsw2  
zig div WD P-fin arm64 thru 38.09 libbid 40.37 1.06× Rzgsw2  
zig div ET P-fin arm64 thru 7.59 libbid 6.09 0.80× Rzgsw2  
zig div PT P-fin arm64 thru 4.13 libbid 6.09 1.47× Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ns alt alt ns ratio run notes
zig add MIX P-fin x86_64 thru 7.37 libbid 30.56 4.15× xRzgsw2  
zig sub MIX P-fin x86_64 thru 8.62 libbid 31.87 3.70× xRzgsw2  
zig mul CP P-fin x86_64 thru 5.96 libbid 50.91 8.54× xRzgsw2  
zig mul WP P-fin x86_64 thru 31.35 libbid 66.63 2.13× xRzgsw2  
zig div CD P-fin x86_64 thru 69.14 libbid 82.11 1.19× xRzgsw2  
zig div WD P-fin x86_64 thru 97.47 libbid 86.90 0.89× xRzgsw2  
zig div ET P-fin x86_64 thru 21.77 libbid 21.25 0.98× xRzgsw2  
zig div PT P-fin x86_64 thru 12.99 libbid 21.42 1.65× xRzgsw2  

Add — SQ · NQ · MQ · OQ · FQ

Swept 4096-input average per band (bare thru; ns/op = Time/4096 over the shared decimal128-resources/swept/ corpus, byte-identical operands every port). Relational peer table with explicit ratios; the cross-port d128 band-shape matrices are in Port-Comparison Benchmark Results.

arm64 (M3 Pro).

port op cat profile arch mode ours alt alt ns ratio run notes
zig add SQss P-gen arm64 thru 1.27 libbid 7.96 6.27× Rzgsw2  
zig add SQos P-gen arm64 thru 2.26 libbid 8.69 3.85× Rzgsw2  
zig add NQss P-gen arm64 thru 4.25 libbid 9.36 2.20× Rzgsw2  
zig add NQos P-gen arm64 thru 4.88 libbid 9.78 2.00× Rzgsw2  
zig add MQss P-gen arm64 thru 6.21 libbid 9.75 1.57× Rzgsw2  
zig add MQos P-gen arm64 thru 12.26 libbid 9.71 0.79× Rzgsw2  
zig add OQss P-gen arm64 thru 12.00 libbid 13.66 1.14× Rzgsw2  
zig add OQos P-gen arm64 thru 15.68 libbid 15.31 0.98× Rzgsw2  
zig add FQss P-gen arm64 thru 8.44 libbid 9.32 1.10× Rzgsw2  
zig add FQos P-gen arm64 thru 10.88 libbid 10.38 0.95× Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ours alt alt ns ratio run notes
zig add SQss P-gen x86_64 thru 5.47 libbid 30.12 5.51× xRzgsw2  
zig add SQos P-gen x86_64 thru 7.97 libbid 27.81 3.49× xRzgsw2  
zig add NQss P-gen x86_64 thru 13.50 libbid 31.74 2.35× xRzgsw2  
zig add NQos P-gen x86_64 thru 14.78 libbid 30.17 2.04× xRzgsw2  
zig add MQss P-gen x86_64 thru 15.48 libbid 29.55 1.91× xRzgsw2  
zig add MQos P-gen x86_64 thru 25.00 libbid 28.78 1.15× xRzgsw2  
zig add OQss P-gen x86_64 thru 30.44 libbid 46.76 1.54× xRzgsw2  
zig add OQos P-gen x86_64 thru 37.51 libbid 47.44 1.26× xRzgsw2  
zig add FQss P-gen x86_64 thru 18.61 libbid 33.02 1.77× xRzgsw2  
zig add FQos P-gen x86_64 thru 23.80 libbid 32.54 1.37× xRzgsw2  

Subtract — SQ · NQ · MQ · OQ · FQ

Swept 4096-input average per band (bare thru; ns/op = Time/4096 over the shared decimal128-resources/swept/ corpus, byte-identical operands every port). Relational peer table with explicit ratios; the cross-port d128 band-shape matrices are in Port-Comparison Benchmark Results.

arm64 (M3 Pro).

port op cat profile arch mode ours alt alt ns ratio run notes
zig sub SQss P-gen arm64 thru 1.69 libbid 9.68 5.73× Rzgsw2  
zig sub SQos P-gen arm64 thru 1.32 libbid 10.04 7.61× Rzgsw2  
zig sub NQss P-gen arm64 thru 4.93 libbid 10.87 2.20× Rzgsw2  
zig sub NQos P-gen arm64 thru 4.29 libbid 11.42 2.66× Rzgsw2  
zig sub MQss P-gen arm64 thru 12.78 libbid 9.94 0.78× Rzgsw2  
zig sub MQos P-gen arm64 thru 6.23 libbid 9.78 1.57× Rzgsw2  
zig sub OQss P-gen arm64 thru 16.28 libbid 15.79 0.97× Rzgsw2  
zig sub OQos P-gen arm64 thru 12.95 libbid 14.84 1.15× Rzgsw2  
zig sub FQss P-gen arm64 thru 11.10 libbid 9.39 0.85× Rzgsw2  
zig sub FQos P-gen arm64 thru 9.44 libbid 9.48 1.00× Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ours alt alt ns ratio run notes
zig sub SQss P-gen x86_64 thru 7.50 libbid 33.73 4.50× xRzgsw2  
zig sub SQos P-gen x86_64 thru 5.48 libbid 32.99 6.02× xRzgsw2  
zig sub NQss P-gen x86_64 thru 14.93 libbid 34.85 2.33× xRzgsw2  
zig sub NQos P-gen x86_64 thru 13.86 libbid 35.61 2.57× xRzgsw2  
zig sub MQss P-gen x86_64 thru 25.01 libbid 34.98 1.40× xRzgsw2  
zig sub MQos P-gen x86_64 thru 15.99 libbid 34.66 2.17× xRzgsw2  
zig sub OQss P-gen x86_64 thru 38.38 libbid 51.72 1.35× xRzgsw2  
zig sub OQos P-gen x86_64 thru 31.54 libbid 50.81 1.61× xRzgsw2  
zig sub FQss P-gen x86_64 thru 23.74 libbid 37.25 1.57× xRzgsw2  
zig sub FQos P-gen x86_64 thru 20.71 libbid 36.53 1.76× xRzgsw2  

Multiply — CP · WP · XP

Swept 4096-input average per band (bare thru; ns/op = Time/4096 over the shared decimal128-resources/swept/ corpus, byte-identical operands every port). Relational peer table with explicit ratios; the cross-port d128 band-shape matrices are in Port-Comparison Benchmark Results.

arm64 (M3 Pro).

port op cat profile arch mode ours alt alt ns ratio run notes
zig mul CP P-gen arm64 thru 1.58 libbid 23.13 14.64× Rzgsw2  
zig mul WP P-gen arm64 thru 18.69 libbid 35.06 1.88× Rzgsw2  
zig mul XP P-gen arm64 thru 22.90 libbid 45.24 1.98× Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ours alt alt ns ratio run notes
zig mul CP P-gen x86_64 thru 7.46 libbid 50.55 6.78× xRzgsw2  
zig mul WP P-gen x86_64 thru 36.01 libbid 73.62 2.04× xRzgsw2  
zig mul XP P-gen x86_64 thru 48.32 libbid 104.96 2.17× xRzgsw2  

Divide — CD · WD · XD (+ ET · PT)

Swept 4096-input average per band (bare thru; ns/op = Time/4096 over the shared decimal128-resources/swept/ corpus, byte-identical operands every port). Relational peer table with explicit ratios; the cross-port d128 band-shape matrices are in Port-Comparison Benchmark Results.

arm64 (M3 Pro).

port op cat profile arch mode ours alt alt ns ratio run notes
zig div CD P-gen arm64 thru 30.28 libbid 36.77 1.21× Rzgsw2  
zig div WD P-gen arm64 thru 40.30 libbid 37.54 0.93× Rzgsw2  
zig div XD P-gen arm64 thru 32.08 libbid 38.97 1.21× Rzgsw2  
zig div ET P-gen arm64 thru 10.70 libbid 11.68 1.09× Rzgsw2  
zig div PT P-gen arm64 thru 4.13 libbid 11.45 2.77× Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ours alt alt ns ratio run notes
zig div CD P-gen x86_64 thru 79.70 libbid 86.65 1.09× xRzgsw2  
zig div WD P-gen x86_64 thru 95.55 libbid 88.12 0.92× xRzgsw2  
zig div XD P-gen x86_64 thru 80.90 libbid 88.89 1.10× xRzgsw2  
zig div ET P-gen x86_64 thru 33.46 libbid 31.98 0.96× xRzgsw2  
zig div PT P-gen x86_64 thru 12.95 libbid 32.28 2.49× xRzgsw2  

FMA — FN (Barrett) · FF (fits-128)

Swept 4096-input average per band (bare thru; ns/op = Time/4096 over the shared decimal128-resources/swept/ corpus, byte-identical operands every port). Relational peer table with explicit ratios; the cross-port d128 band-shape matrices are in Port-Comparison Benchmark Results.

arm64 (M3 Pro).

port op cat profile arch mode ours alt alt ns ratio run notes
zig fma FN FMA arm64 thru 64.07 - - - Rzgsw2  
zig fma FF FMA arm64 thru 41.72 - - - Rzgsw2  

x86_64 (Intel i9-9880H).

port op cat profile arch mode ours alt alt ns ratio run notes
zig fma FN FMA x86_64 thru 119.12 - - - xRzgsw2  
zig fma FF FMA x86_64 thru 79.83 - - - xRzgsw2