Floating Point to Fixed and Fixed to Floating Point Conversion by sweeksBigBlue · Pull Request #242 · intel/rohd-hcl

sweeksBigBlue · 2025-07-17T20:09:18Z

This pull request is a base line implementation of PR 112.

In this PR we make a generalized converter for Fixed to Float and Float to Fixed method implementations. This should make it easier in the future for designers to implement transitions between the two elements and expand hardware design capabilities.

sweeksBigBlue · 2025-07-17T20:09:40Z

@desmonddak and @mkorbel1 we are finally ready to review this!

Signed-off-by: James Farwell <james.c.farwell@intel.com>

mkorbel1

Looking very good on the right track!

lib/src/arithmetic/values/fixed_point_value.dart

doc/components/fixed_point.md

mkorbel1 · 2025-07-17T20:16:38Z

lib/src/arithmetic/values/floating_point_values/floating_point_value.dart

  }
+
+  /// Losslessly convert a [FloatingPointValue] to a [FixedPointValue].
+  /// TODO(jcfarwe):


Did you mean to include these TODO's in this PR or do them?

Also, TODOs should be after // rather than /// (since they are not part of the doc comment)

mkorbel1 · 2025-07-17T20:17:48Z

test/arithmetic/values/fixed_point_value_test.dart

      }
    }
  });
+


what about testing with vs without explicitJBit?

Do you mind expanding on that? Idk what that is but happy to add it to my tests.

https://github.com/intel/rohd-hcl/blob/main/doc/components/floating_point.md#explicit-j-bit

I'll test the documentation on you: if this doc doesn't explain it well enough, we should improve it!

Great question! I think maybe this would be an excellent second issue I can follow-up on in a different PR maybe? Honest answer is, even with reading that I'd need time to digest that and maybe that should be an expanded method to this base one? Or maybe I'm over complicating that.

I think it should be a relatively straight-forward change, basically even normal things look like sub-normals with an explicit 1. @desmonddak what do you think?

Sorry for the late reply: it should be quite simple; a small change in how you adjust the mantissa in each direction.
So going from fixed value to normal float value: you would traditionally compute the mantissa for the floating point value and chop off the leading '1' and store the rest of the mantissa into the FPV.

But going to an explicit j-bit form, you would keep that leading 1 and put it into the mantissa. The FPV feels effectively one bit shorter because you need to store the leading '1. Note that you would do this just as you do with subnormals. In other words, if you fixed point was small enough to require a subnormal representation in FPV, you would also keep that leading '1'.

Going from an explicitJBit normal float to fixed, you would recognize that you don't need to prefix the mantissa with a '1' representing the j-bit. The leading '1' is already in the mantissa. This is just like it is for subnormal. If you are converting a subnormal float to fixed, you do not prefix the mantissa with a '1'.

To test, you need to provide an explicitJBit form of floatingPointValue to translate to fixed and conversely, indicate that you want an explicitJBit form of floatingPointValue when converting from fixed.

Explicit JBIT is experimental: we recognize that we do not have to normalize in some situations as it doesn't produce any more accuracy (say adding two narrow mantissa FPs to get a wider mantissa FP). Rather than normalizing, we can leave it in explicit j-bit form for the next operation and save a normalization step, and let normalization happen after the next operation.

Something like this;
test('FixedPointValue: toFloatingPointValue signed', () {
const m = 10;
const n = 12;
const width = m + n + 1; // 1 for sign bit
for (var i = 0; i < pow(2, width); i++) {
final testVal = (i / pow(2, width)).toInt();
final fxv = FixedPointValue.populator(
integerWidth: m, fractionWidth: n, signed: true)
.ofLogicValue(LogicValue.ofInt(testVal, width));
FloatingPointValue fpv;
if (testVal > 0) {
fpv = fxv.toFloatingPointValue(false); // explicitJBit
} else {
fpv = fxv.toFloatingPointValue(true);
}
expect(fpv.toDouble(), fxv.toDouble(),
reason: 'toFloatingPointValue failed for $testVal');
}
});
?

@jcfarwe has pushed tests for J-bit, ready for review @desmonddak and @mkorbel1

Tests look good. Can you merge the signed/unsigned versions into one?

Can you shorten the widths to speed up the tests? 22 bits is quite wide.

I tried the following which ran in about 1s:

for (final (m, n) in [(5, 7), (7, 5), (0, 7), (7, 0)]) {

lib/src/arithmetic/values/fixed_point_value.dart

mkorbel1 · 2025-07-17T20:23:12Z

test/arithmetic/values/floating_point_value_test.dart

  });
+  test('FPV: toFixedPointValue', () async {
+    //
+    // TODO(jcfarwe): Find all corner cases, automate test cases.


lingering TODO?

mkorbel1 · 2025-07-17T20:24:01Z

test/arithmetic/values/floating_point_value_test.dart

      }
    });
  });
+  test('FPV: toFixedPointValue', () async {


maybe add an exhaustive run as well for at least some width(s)?

Added exhaustive test for various exponent and mantissa widths

desmonddak · 2025-07-17T20:56:09Z

lib/src/arithmetic/values/fixed_point_value.dart


-  /// Addition operation that returns a [FixedPointValue].
+  /// Converts a [FixedPointValue] to a [FloatingPointValue].
+  FloatingPointValue toFloat() {


Should this be toFloatingPointValue()?

Great point, I can fix that if you think that is more appropriate

Updated name, it is more clear and obvious good catch.

test/arithmetic/values/fixed_point_value_test.dart

desmonddak · 2025-07-17T21:19:03Z

test/arithmetic/values/floating_point_value_test.dart

+    ];
+
+    for (final testCase in testCases) {
+      final bias = (pow(2, testCase[1] - 1) - 1).toInt();


Removing the sizes and using the FPV populator might be more readable for this test.
It's a bit confusing whether the sizes are for the FPV or the FXV.

See compound_adder_test.dart:262 for an example of list structuring and using Boolean literals. It's not great but it helps to structure the list to help the reader parse (the structure has separated inputs and outputs, for example).

desmonddak · 2025-07-17T21:25:37Z

lib/src/arithmetic/values/fixed_point_value.dart

+          mantissa: LogicValue.zero);
+    }
+
+    // FOR loop to find first one to figure out correct exponent value and width


On comments: this can confuse people. Find the first 1 .would be better, or 'Count the leading zeros in order to correct the exponent value.' may be better still. No need to say 'FOR loop'. :j)

If you like to use one, then use it like this one and zero.

Remember to end on a period (Dart conventions).

Fixed comments, good call :)

Signed-off-by: James Farwell <james.c.farwell@intel.com>

desmonddak · 2025-07-28T17:46:39Z

test/arithmetic/values/floating_point_value_test.dart

+          for (var testExp = minExp; testExp < maxExp; testExp++) {
+            final bias = (pow(2, expWidth - 1) - 1).toInt();
+            final exp = LogicValue.ofInt(testExp, expWidth);
+            final sign = LogicValue.ofInt(0, 1);


Do you also want to put 'sign in a loop:

for (final sign in [LogicValue.zero, Logicvalue.one]) {

Signed-off-by: James Farwell <james.c.farwell@intel.com>

sweeksBigBlue · 2025-07-31T20:41:54Z

@mkorbel1 and @desmonddak just back from vacation, where are we at with the PR? What changes are still outstanding that you feel needs addressed?

desmonddak · 2025-08-25T16:28:04Z

test/arithmetic/values/fixed_point_value_test.dart

+  });
+
+  test('FixedPointValue: toFloatingPointValue unsigned', () {
+    for (final (m, n) in [(10, 12), (0, 22), (22, 0)]) {


Definitely narrow these widths...

desmonddak · 2025-08-25T16:29:06Z

test/arithmetic/values/floating_point_value_test.dart

@@ -631,4 +631,152 @@ void main() {
      }
    });
  });


We like to see a blank line between tests or groups.

desmonddak · 2025-08-25T16:33:10Z

test/arithmetic/values/floating_point_value_test.dart

+      ).ofLogicValue(finalMantissa);
+    }
+
+    test('FPV: toFixedPointValue exhaustive', () async {


Something is fishy about this test. It only fills the upper 20 bits or so of the integer portion, yet the test is labeled 'exhaustive'.

Here is what I see when I print fxv:

(0 0010000000000000000101101100000000000000000000000000000000000000000000000000000000000000000000 00000000000000000000000) (1 1101111111111111111010010100000000000000000000000000000000000000000000000000000000000000000000 00000000000000000000000)

Updating simple fixed-to-float and float-to-fixed functionality to ROHD

e9ed100

Signed-off-by: James Farwell <james.c.farwell@intel.com>

sweeksBigBlue force-pushed the fp2fpcvt branch from f2c376d to e9ed100 Compare July 17, 2025 20:20

Fixing merge issue

62dd49b

mkorbel1 requested changes Jul 17, 2025

View reviewed changes

sweeksBigBlue added 3 commits July 17, 2025 20:28

Adding back in comment for fixed point

fe901ab

Removing redundant line

ae26223

Fixing doc wording

c0fbb9c

desmonddak reviewed Jul 17, 2025

View reviewed changes

test/arithmetic/values/fixed_point_value_test.dart Show resolved Hide resolved

Updating name to toFloatingPointValue()

ca9ac13

desmonddak reviewed Jul 17, 2025

View reviewed changes

jcfarwe and others added 8 commits July 21, 2025 16:14

Add exhaustive test

8ab25ea

Signed-off-by: James Farwell <james.c.farwell@intel.com>

add explicitJBit test

552a1c3

Signed-off-by: James Farwell <james.c.farwell@intel.com>

fix explicitJBit functionality

b1f3250

Signed-off-by: James Farwell <james.c.farwell@intel.com>

Cleaning up j-bit testing

fc6d0f2

Fixing comments and updating edge cases to fixed point test

b41af80

Adding error catch for zero width

5ab6019

Change test cases from list to tuple, name each param

c4c4eeb

Signed-off-by: James Farwell <james.c.farwell@intel.com>

fix linter issues

6c94c7f

Signed-off-by: James Farwell <james.c.farwell@intel.com>

desmonddak reviewed Jul 28, 2025

View reviewed changes

add +/- sign to exhaustive test

ce13380

Signed-off-by: James Farwell <james.c.farwell@intel.com>

desmonddak reviewed Aug 25, 2025

View reviewed changes

@@ @@ -631,4 +631,152 @@ void main() { @@
                     }
                   });
                 });

Conversation

sweeksBigBlue commented Jul 17, 2025

Uh oh!

sweeksBigBlue commented Jul 17, 2025

Uh oh!

mkorbel1 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

desmonddak Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

desmonddak Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sweeksBigBlue commented Jul 31, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

desmonddak Aug 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

desmonddak Aug 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

desmonddak Jul 17, 2025 •

edited

Loading

desmonddak Jul 17, 2025 •

edited

Loading

desmonddak Aug 25, 2025 •

edited

Loading

desmonddak Aug 25, 2025 •

edited

Loading