Floating

[desmonddak]

Description & Motivation

This PR is focused on resolving issues reported as well as TODOs documented on the arithmetic library portion of ROHD-HCL. The biggest fixes are around rounding and conversions.

Related Issue(s)

Features

• Added IEEE rounding modes across floating-point values and logic operations.
  Fixes Add all rounding modes #191.
• Unified fixed- and floating-point value rounding APIs. Fixes Unify rounding modes #173.
• Added exact FloatingPointValue/FixedPointValue conversions. Fixes Add converters betwee FloatingPointValue and FixedPointValue #112.
• Added toLogic() conversion for fixed- and floating-point values. Fixes Add an easy way for arithmetic Values to convert to Const versions of corresponding signals #200.
• Added typed fixed- and floating-point arithmetic signal operations. Fixes Add basic operations directly on arithmetic types #199.
• Pushed runtime configuration control through multipliers based on PR pushed dynamic configuration through all multiplier componentry #222.

Fixes

• Corrected floating-point multiplier rounding, narrow-output behavior, and
  zero multiplication. Fixes FloatingPointMultiplier to narrower FP type #194.
• Corrected round-nearest-even behavior across mantissa widths and subnormal
  boundaries. Fixes RoundRNE is not rounding to even, both logic and value side #190.
• Corrected subnormal FloatingPointConverter conversions and added exhaustive
  implicit-j-bit subnormal coverage. Fixes FloatingPointConverter fails on narrowing mantissa with explicitJBit #241.
• Corrected IEEE-754 NaN comparisons and encodings. Fixes FP NaN are not comparable (e.g. nan != nan) #252 and Fixes Maximum Exponent for E4M3 #115.
• Corrected FloatingPointValue.ulp() for subnormals. Fixes FloatingPointValue unit of least precision (ulp) is incorrect for subnormals #206.
• Made FloatingPointValue.operator / exact and correctly rounded. Fixes Make FloatingPointValue operations scale beyond double #113.
• Corrected FloatToFixed explicit-j-bit conversion, overflow handling, and
  discarded-bit sticky rounding.
• Added early validation for multiply-accumulate output widths and runtime
  boolean control widths.

API Cleanup and Verification

• Added StaticOrRuntimeControl<T> as the general static-or-runtime control
  abstraction; deprecated BooleanConfig and RuntimeConfig.
• Expanded exhaustive conversion, rounding, arithmetic, and
  multiply-accumulate coverage.

Testing

All existing tests pass. Several new tests added to cover bugs reported and fixed.

Backwards-compatibility

Is this a breaking change that will not be backwards-compatible? If yes, how so?

• FixedPointValue, its factory, and its populator now default to
  signed: true, matching FixedPoint. Pass signed: false to retain the
  previous unsigned behavior. Fixes FixedPoint and FixedPointValue have different default values for signed #249.

Documentation

Does the change require any updates to documentation? If so, where? Are they included?

Yes. Relevant documentation was updated to match the new functionality introduced (e.g., StaticOrRuntimeControl.

[desmonddak]

@copilot resolve the merge conflicts in this pull request

[Copilot]

🟡 Changes recommended

Mixed-sign MAC handling, IEEE status flags, converter sticky rounding, and subnormal square roots contain correctness issues.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Expands ROHD-HCL arithmetic with IEEE rounding, exact conversions, typed operations, configurable multiply-accumulate logic, and static/runtime controls.

Changes:

• Adds rounding modes, floating-point status flags, NaN handling, and conversion fixes.
• Adds typed fixed/floating arithmetic and configurable MAC behavior.
• Expands tests, compatibility APIs, documentation, and release notes.

File summaries

File	Description
test/summation/counter_test.dart	Updates copyright.
test/static_or_runtime_control_test.dart	Tests legacy import compatibility.
test/memory/resettable_entries_test.dart	Tests runtime reset values.
test/arithmetic/sign_magnitude_adder_test.dart	Tests carry-in behavior.
test/arithmetic/ones_complement_adder_test.dart	Migrates subtraction control API.
test/arithmetic/multiplier_test.dart	Expands MAC width and control tests.
test/arithmetic/multiplier_encoder_test.dart	Reduces test widths and cleanup.
test/arithmetic/generic_multiply_accumulate_test.dart	Tests generic MAC functionality.
test/arithmetic/floating_point/floating_point_test.dart	Tests NaNs, comparisons, and operators.
test/arithmetic/floating_point/floating_point_status_test.dart	Tests IEEE exception status.
test/arithmetic/floating_point/floating_point_sqrt_test.dart	Tests square-root rounding and subnormals.
test/arithmetic/floating_point/floating_point_rounding_test.dart	Tests all rounding modes.
test/arithmetic/floating_point/floating_point_conversion_test.dart	Expands exhaustive conversion coverage.
test/arithmetic/floating_point/floating_point_adder_test.dart	Uses exact value-side addition.
test/arithmetic/floating_point/floating_point_adder_dualpath_test.dart	Expands dual-path rounding tests.
test/arithmetic/float_to_fixed_test.dart	Tests exact and rounded conversions.
test/arithmetic/fixed_to_float_test.dart	Tests rounding and explicit J bits.
test/arithmetic/fixed_point_test.dart	Tests typed arithmetic and constants.
test/arithmetic/dotproduct_test.dart	Tests dynamic partial-sum extension.
test/arithmetic/compound_adder_test.dart	Migrates subtraction controls.
test/arithmetic/adder_test.dart	Migrates subtraction controls.
test/apb_bfm_test.dart	Updates copyright.
lib/src/static_or_runtime_parameter.dart	Preserves the legacy import path.
lib/src/static_or_runtime_control.dart	Adds generic static/runtime controls.
lib/src/models/memory_model.dart	Updates copyright.
lib/src/arithmetic/values/floating_point_values/floating_point_rounding_mode.dart	Adds exact BigInt rounding.
lib/src/arithmetic/values/floating_point_values/floating_point_8_e4m3_value.dart	Corrects E4M3 limits and NaNs.
lib/src/arithmetic/values/fixed_point_value.dart	Adds signed APIs and exact conversions.
lib/src/arithmetic/values/fixed_point_populator.dart	Adds rounded exact population APIs.
lib/src/arithmetic/signals/floating_point_logics/floating_point_logic.dart	Adds constants, arithmetic, and IEEE comparisons.
lib/src/arithmetic/signals/fixed_point_logic.dart	Adds typed fixed-point arithmetic.
lib/src/arithmetic/sign_magnitude_adder.dart	Adds carry-in support.
lib/src/arithmetic/ones_complement_adder.dart	Unifies static/runtime subtraction.
lib/src/arithmetic/multiply_accumulate.dart	Adds output widths and generic MAC.
lib/src/arithmetic/multiplier_components/partial_product_sign_extend.dart	Narrows unused sign-extension logic.
lib/src/arithmetic/floating_point/floating_point.dart	Exports status support.
lib/src/arithmetic/floating_point/floating_point_status.dart	Defines IEEE status signals.
lib/src/arithmetic/floating_point/floating_point_sqrt.dart	Adds rounding and status interfaces.
lib/src/arithmetic/floating_point/floating_point_sqrt_simple.dart	Adds rounded subnormal-capable square root.
lib/src/arithmetic/floating_point/floating_point_rounding.dart	Implements shared hardware rounders.
lib/src/arithmetic/floating_point/floating_point_multiplier.dart	Adds multiplier status output.
lib/src/arithmetic/floating_point/floating_point_multiplier_simple.dart	Adds narrowing, rounding, and status.
lib/src/arithmetic/floating_point/floating_point_converter.dart	Adds rounded conversion and status.
lib/src/arithmetic/floating_point/floating_point_adder.dart	Adds adder status output.
lib/src/arithmetic/floating_point/floating_point_adder_singlepath.dart	Expands rounding and special handling.
lib/src/arithmetic/float_to_fixed.dart	Adds rounding and overflow fixes.
lib/src/arithmetic/fixed_to_float.dart	Adds shared rounding and explicit J bits.
lib/src/arithmetic/dotproduct.dart	Corrects reduction widths and extension.
lib/src/arithmetic/compound_adder.dart	Unifies subtraction configuration.
lib/rohd_hcl.dart	Exports the new control API.
doc/static_or_runtime_parameters.md	Documents static/runtime controls.
doc/README.md	Links the new documentation.
doc/components/floating_point.md	Documents floating-point operators.
doc/components/fixed_point.md	Documents signed APIs and arithmetic.
CHANGELOG.md	Records features, fixes, and deprecations.

Review details

• Files reviewed: 62/62 changed files
• Comments generated: 13
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

Square-root handling, directed conversion overflow, MAC definition naming, and metadata consistency have unresolved correctness issues.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (3)

Previously missed (1) — in code that hasn't changed since the last review.

lib/src/arithmetic/fixed_to_float.dart:46

• The new rounding mode affects significand rounding, but overflow still unconditionally emits infinity in the expoMoreThanMax branch. Directed/toward-zero modes must instead return the largest finite value when the rounding direction is away from infinity (for example, +15 converted to E3M2 with roundTowardsZero should become +14, not +infinity).

lib/src/arithmetic/floating_point/floating_point_sqrt_simple.dart:199

• Underflow is hard-wired low even when the newly supported square root produces an inexact subnormal result. This contradicts FloatingPointStatus.underflow's “tiny result was inexact after rounding” contract; the smallest E3M5 subnormal exercised by the new test is one such case.

    internalStatus.underflow <= Const(0);

lib/src/arithmetic/values/fixed_point_value.dart:52

• The PR's backwards-compatibility section says FixedPointValue now defaults to signed: true, but this constructor deliberately retains false (as does the deprecated populator), while the changelog describes that preservation. Align the PR description with this compatibility strategy, or change the implementation if the stated breaking default is intended.

  factory FixedPointValue(
          {required LogicValue integer,
          required LogicValue fraction,
          bool signed = false}) =>
      FixedPointValue.withSignedness(
          integer: integer, fraction: fraction, signed: signed);

• Files reviewed: 62/62 changed files
• Comments generated: 3
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Mixed-signed MAC arithmetic, conversion overflow rounding, and square-root underflow status contain unresolved correctness issues.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (3)

lib/src/arithmetic/float_to_fixed.dart:173

• Overflow is calculated from the unrounded magnitude and never includes rounder.doRound. For signed Q2.1, E4M4 input 3.625 with round-nearest-even truncates to raw 7 (apparently fitting), then rounds to raw 8, wraps to -4, and leaves overflow false. Compute overflow from the rounded magnitude so rounding carry across the positive limit is reported.

          overflow <
              (_signedGtConst(shift, threshold) |
                  (shift.eq(Const(threshold, width: shift.width)) &
                      ~(float.sign & atThresholdIsExactPowerOfTwo))),

lib/src/arithmetic/values/fixed_point_value.dart:52

• The PR's backwards-compatibility section says FixedPointValue and its factory now default to signed: true, but this public factory still defaults to false and forwards that value. This also leaves the API mismatch from #249 in place for existing callers; either implement the documented breaking default or update the PR description to state that only the new withSignedness API uses the signed default.

  factory FixedPointValue(
          {required LogicValue integer,
          required LogicValue fraction,
          bool signed = false}) =>
      FixedPointValue.withSignedness(
          integer: integer, fraction: fraction, signed: signed);

lib/src/arithmetic/multiply_accumulate.dart:440

• commonWidth + 1 is insufficient when exactly one operand is signed. For example, the new test computes unsigned 49 + signed 15: the exact result is 64, but in 7 bits bit 6 is mistaken for a sign bit and widening produces 192. Reserve two bits beyond the widest operand so every mixed-signed sum remains representable before fitAccumulateWidth.

    final commonWidth = (product.width < c.width ? c.width : product.width) + 1;

• Files reviewed: 62/62 changed files
• Comments generated: 3
• Review effort level: Balanced

[Copilot]

🟡 Changes recommended

Explicit-J encodings, conversion rounding, pipelined signedness, and generated definition naming contain unresolved correctness issues.

Get a fresh assessment by requesting another Copilot review.

Review details

Suppressed comments (4)

Previously missed (1) — in code that hasn't changed since the last review.

lib/src/arithmetic/floating_point/floating_point_converter.dart:190

• The rounding branch is selected by stored mantissa width, but explicit-J formats spend one stored bit on the J bit. An implicit M4 source converted to an explicit M4 destination loses one fraction bit, yet this condition skips the rounder, so all rounding modes behave like truncation (for example, positive 1.0625 should become 1.125 when rounding toward +infinity). Compare effective fraction widths and adjust the retained windows accordingly.

lib/src/arithmetic/signals/floating_point_logics/floating_point_logic.dart:315

• For explicit-J-bit formats this sets only the J bit, leaving every fraction bit zero. The resulting value is classified as infinity rather than NaN (isNaN checks the fractional portion), so FloatingPoint.nan(explicitJBit: true, ...) is broken. Build the constant through the value populator so the J bit and quiet-NaN bit are both encoded correctly.

    final mantissa =
        Const(BigInt.one << (mantissaWidth - 1), width: mantissaWidth);

lib/src/arithmetic/values/fixed_point_value.dart:52

• The PR description says the FixedPointValue factory now defaults to signed: true, but this public factory still defaults to false (as does the deprecated populator). The code and changelog preserve backward compatibility, so either update the PR compatibility statement or change these defaults to match the stated breaking change.

  factory FixedPointValue(
          {required LogicValue integer,
          required LogicValue fraction,
          bool signed = false}) =>
      FixedPointValue.withSignedness(
          integer: integer, fraction: fraction, signed: signed);

lib/src/arithmetic/fixed_to_float.dart:242

• For E4M3, expoMoreThanMax is still computed against encoded exponent 14 even though exponent 15 contains finite values. This branch therefore maps representable values such as 256 to the largest finite value (448). Derive the overflow limit from largestFinite.exponent rather than the generic infinity-reserving limit.

        ElseIf(expoMoreThanMax, [
          _convertedFloat.exponent < Const(largestFinite.exponent),
          _convertedFloat.mantissa < Const(largestFinite.mantissa),

• Files reviewed: 62/62 changed files
• Comments generated: 6
• Review effort level: Balanced

[Copilot]

Copilot review overview

🟡 Changes recommended

Format-specific conversions and square-root boundary handling currently produce incorrect floating-point encodings.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 3 · 3

Open (6)

• Allow representable E4M3 exponent-15 finite values · New
• Compare overflow against destination finite limits · New
• Preserve concrete floating-point format in constant conversion · New
• A NaN payload can contain discarded bits, making roundingInexact true even though the converter…
• This sticky calculation only notices a right shift when every set bit is lost. If a shift discards…
• roundingInexact is derived from the arithmetic datapath even when the selected result is an…

Resolved since last review (6)

• When clk is present, adder.sum is delayed by the compressor pipeline, but fitAccumulateWidth…
• Now that the multiplier generates different rounding circuits, its automatic definition name must…
• This newly mode-dependent rounder is emitted under an automatic adder definition name that does not…
• The selected rounding mode affects this module's hardware, but it is absent from the automatic…
• roundingMode now changes the generated circuit, but the automatic definition name is unchanged.…
• This infinity encoding writes an all-zero mantissa. For an explicit-J-bit destination, infinity…

Previously missed (1)

In code that hasn't changed since last review

Handle zero biased exponents as subnormal results

lib/​src/​arithmetic/​floating_point/​floating_point_sqrt_simple.dart:152

Biased exponent zero is also a subnormal result, but this condition only detects negative roundedExponent. For example, sqrt(E3M5 raw 0 000 00010) is exactly raw 0 000 10000; here roundedExponent is zero, so the current normal path emits exponent zero with the hidden bit discarded (zero instead of 0.125). Include zero in the subnormal condition.

[Copilot]

Copilot review overview

🟡 Changes recommended

Several generated-definition collisions and explicit-J-bit encoding failures remain unresolved.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 5 · 1

Open (6)

• Include explicit-J mode in float definition names · New
• Include FloatToFixed configuration in definition names · New
• Include source and destination J-bit modes in converter names · New
• Include generators and signedness in MAC definition names · New
• Avoid zero-width payload signals during NaN conversion · New
• roundingInexact is derived from the arithmetic datapath even when the selected result is an…

Resolved since last review (5)

• Preserve concrete floating-point format in constant conversion
• Compare overflow against destination finite limits
• Allow representable E4M3 exponent-15 finite values
• A NaN payload can contain discarded bits, making roundingInexact true even though the converter…
• This sticky calculation only notices a right shift when every set bit is lost. If a shift discards…

[Copilot]

Copilot review overview

🟡 Changes recommended

Explicit-J overflow encoding, HDL definition-name collisions, and API contract issues remain unresolved.

Review effort: Balanced
Findings: 2 · 2

Open (4)

• Include adder generator identity in compression-tree definition name · New
• Include adder generator identity in reduction definition name · New
• Avoid double-width increase for unsigned subtraction · New
• Make pipelined signedness test observe registered sign extension · New

Resolved since last review (6)

• Avoid zero-width payload signals during NaN conversion
• Include generators and signedness in MAC definition names
• Include source and destination J-bit modes in converter names
• Include FloatToFixed configuration in definition names
• Include explicit-J mode in float definition names
• roundingInexact is derived from the arithmetic datapath even when the selected result is an…

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.