srielau commented on PR #58410:
URL: https://github.com/apache/spark/pull/58410#issuecomment-5463526152
# LibBID
LibBID implements IEEE 754 decimal64 and decimal128 arithmetic using the
Binary Integer Decimal (BID) encoding.
## Performance compared with `BigDecimal`
The following results compare the `Bid64` object API with `BigDecimal` under
`MathContext.DECIMAL64`, and the `Bid128` object API with `BigDecimal` under
`MathContext.DECIMAL128`. All values are average latency in nanoseconds per
operation. The ratio is `BigDecimal / BID`, so a value greater than 1 means
that BID completed the benchmark faster.
| Function / workload | BID64 | BigDecimal DECIMAL64 | Ratio | BID128 |
BigDecimal DECIMAL128 | Ratio |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| add (same quantum) | 4.69 | 21.03 | 4.48x | 11.60 | 17.28 | 1.49x |
| add (mixed quantum) | 10.55 | 38.94 | 3.69x | 88.86 | 92.54 | 1.04x |
| add (full precision) | 9.64 | 43.73 | 4.54x | 67.14 | 113.85 | 1.70x |
| subtract | 4.94 | 25.76 | 5.22x | 11.58 | 26.34 | 2.27x |
| multiply (same quantum) | 17.88 | 23.47 | 1.31x | 69.44 | 158.61 | 2.28x |
| multiply (mixed quantum) | 10.54 | 10.72 | 1.02x | 31.88 | 62.76 | 1.97x |
| multiply (full precision) | 22.78 | 29.08 | 1.28x | 103.24 | 181.58 |
1.76x |
| divide (same quantum) | 20.24 | 27.52 | 1.36x | 161.05 | 155.89 | 0.97x |
| divide (mixed quantum) | 22.23 | 40.01 | 1.80x | 142.83 | 187.67 | 1.31x |
| divide (full precision) | 23.88 | 27.54 | 1.15x | 168.53 | 179.87 | 1.07x |
| square root | 154.30 | 1287.99 | 8.35x | 864.41 | 2639.21 | 3.05x |
| fused multiply-add | 119.67 | 136.99 | 1.14x | 195.11 | 247.74 | 1.27x |
| truncating remainder (`fmod`) | 10.03 | 174.07 | 17.36x | 278.49 | 327.17
| 1.17x |
| positive integral power | 12.92 | 78.52 | 6.08x | 37.88 | 111.36 | 2.94x |
| round to integral | 7.16 | 11.58 | 1.62x | 17.84 | 76.46 | 4.28x |
| scale by power of ten | 3.61 | 3.52 | 0.98x | 3.90 | 3.69 | 0.95x |
| ordered less-than | 5.90 | 4.43 | 0.75x | 10.01 | 4.59 | 0.46x |
| cohort-equal numeric comparison | 8.43 | 7.53 | 0.89x | 14.92 | 19.42 |
1.30x |
The add, multiply, and divide workloads are:
- **same quantum:** operands have similar precision and the same exponent;
- **mixed quantum:** operands have varied precision and different exponents;
- **full precision:** operands use the format's maximum precision and have
different exponents.
The remaining arithmetic benchmarks use positive, full-precision values in
`[1, 10)`, signed full-precision FMA addends, integral powers from 2 through
5,
and decimal scale changes from -12 through 12.
The FMA comparison performs an exact `BigDecimal` multiply and add followed
by
one rounding, matching fused semantics. BID `fmod` and
`BigDecimal.remainder` both use a quotient truncated toward zero. The power
comparison covers positive integral exponents. Round-to-integral compares
numeric results; `BigDecimal` does not retain BID cohorts or signed zero.
These results were collected by the full JMH profile on August 29, 2026 from
libbid-java commit `9936d04`, before the source was incorporated into Spark:
- JMH 1.37;
- OpenJDK 17.0.15;
- Intel Xeon 6975P-C;
- one benchmark thread;
- two forks;
- three 1-second warmup iterations per fork;
- five 1-second measurement iterations per fork;
- 1 GiB fixed heap with `AlwaysPreTouch`; and
- operands prepared outside the measured region.
Microbenchmark results are host- and JVM-specific. Comparisons require the
same JVM, host, benchmark inputs, and JMH settings.
## Capabilities not provided by `BigDecimal`
`BigDecimal` is an arbitrary-precision finite decimal type. LibBID implements
the fixed-size IEEE 754 decimal floating-point model, including:
- exact 64-bit decimal64 and 128-bit decimal128 interchange encodings;
- positive and negative zero, positive and negative infinity, quiet NaN,
signaling NaN, normal values, and subnormal values;
- canonicality checks and complete IEEE decimal classification;
- explicit sticky status flags for invalid operation, denormal operand,
division by zero, overflow, underflow, and inexact results;
- IEEE quiet and signaling comparisons, unordered predicates, `totalOrder`,
`totalOrderMag`, cohort equality, and `sameQuantum`;
- cohort- and quantum-aware operations `quantize`, `quantum`, and
`quantumExponent`;
- adjacent-value operations `nextUp`, `nextDown`, and `nextAfter`;
- IEEE `minNum`, `maxNum`, magnitude variants, and positive difference;
- fused multiply-add with one rounding and IEEE status reporting;
- both IEEE remainder and truncating remainder (`fmod`);
- general decimal floating-point `pow` and `hypot`;
- exponentials `exp`, `expm1`, `exp2`, and `exp10`;
- logarithms `log`, `log1p`, `log2`, and `log10`;
- trigonometric functions `sin`, `cos`, `tan`, `asin`, `acos`, `atan`, and
`atan2`;
- hyperbolic functions `sinh`, `cosh`, `tanh`, `asinh`, `acosh`, and `atanh`;
- `cbrt`, `erf`, `erfc`, `tgamma`, and `lgamma`;
- conversion between BID and DPD encodings and between decimal64 and
decimal128; and
- conversions to and from binary32, binary64, and binary128 with explicit
rounding and status reporting.
The library exposes immutable `Bid64` and `Bid128` objects as well as
allocation-conscious `Bid64Raw` and `Bid128Raw` kernels over encoded bits.
Arithmetic and conversion operations report the full IEEE status flags.
Transcendental compatibility specifies invalid-operation and division-by-zero
flags; Intel's reference transcendental vectors do not consistently specify
inexact, overflow, or underflow flags.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
