Boolean OR, AND, XOR, BitCount: Reduction operators over integers

[KrishnaPG]

Please make sure that this is a feature request. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:feature_template

System information

• TensorFlow.js version (you are using): 4.1.0
• Are you willing to contribute it (Yes/No): Would be glad to but do not have the expertise

Describe the feature and the current behavior/state.
Kindly add boolean AND, OR, XOR , bitCount as reduction operators over integer typed tensors.

For example, for a 2-dimensional matrix, column-wise boolean AND/OR/XOR reduction should result in a 1-dimensional tensor etc. Similarly row-wise etc.

Will this change the current api? How?
New reduction operators may need to be added in addition to the existing sum, prod etc. operators.

These below new reduction operators may have to be added for the integer dType:

• boolean reduction operators: AND, OR, XOR
• bitCount reduction operator that returns the total number of bits that are 1 in the given axis

Who will benefit with this feature?

• Boolean operators enable bit-packing, which will vastly reduce the memory requirement in boolean matrix manipulation senarios. For example, boolean matrix multiplication currently forces us to use "int32" dType, (the bool type is mostly restricted to just logical operators currently it seems). Enabling boolean reduction operators allows extending the boolean operations further into arithmetic, than just logical.

Any Other info.

[Linchenn]

Hi @KrishnaPG , do you have use cases for these ops, especially some well-known models?

Have you found similar APIs from TensorFlow's API? I just found TensorFlow's bitwise ops.

[KrishnaPG]

Thank you @Linchenn The tf.bitwise from the link you have suggested is exactly what is needed - but for some reason I could not find the tf.bitwise in the JS version (documentation), though.

Is it available in the JS version internally (but not documented?)

[Linchenn]

No, it has not been implemented in TensorFlow.js. It makes sense to implement this.

Do you have use cases for this, or have you seen any popular or well known models that have these ops? It would be appreciated.

[KrishnaPG]

I am not sure of ML Models that use these, @Linchenn . I was testing the tf for general vector arithmetic over bool (scientific computing) and it became very difficult without the boolean operators (Shift, AND, OR, bitCount ...). Since the bool logical operators were already present (ANY, ALL ...), I felt these other operators were missing and hence thought of asking to see if they can be added.

[Linchenn]

Got it. Thank you for spotting it out!

[lcrh]

This would be great for hashing data, e.g., for bucketing. I'm currently encountering this while trying to work around missing XLA_JIT support for tf.fingerprint.