[DAPHNE-#523] DaphneDSL flexible matrix/frame literals

[AlexRTer]

This PR implements #523 and allows for matrix literal elements to be expressions of different value types instead of only literals with the same value type. Furthermore, matrix literal dimensions can now be set directly using parentheses ([...](dim_x, dim_y)) instead of having to call reshape. Lastly, it adds new syntax to create column-major as well as row-major frames in DaphneDSL. It also resolves #584 and #622.

Features:

• Flexible value types: Matrix and frame literals are allowed to have elements of different value types. These are then casted to the most general value type among them (or within a column for frames). Currently supported types are double, int64_t, and bool (in order of generality).
  • [true, 2, 3.0] becomes a matrix of double type
• Complex expressions for elements: Matrix and frame literal elements can now be expressions. This includes (but is not limited to) built-in as well as user-defined functions, and (prior defined) variables. Their output only needs to be of a supported type (see above). Currently, there is a limitation for row-major frames as explained below.
  • [1, sqrt(9), sum([1, 4, 2] + fill(3, 3, 1))] -> [1, 9, 16]
• Reshaping: Matrix literals can now be reshaped by simply adding parentheses behind the matrix literal with comma-separated expressions for its dimensions. Both dimensions can be complex expressions. Matrices default to column-matrices if no dimension is given or infer missing dimensions if only 1 has been given.
  • [1, 2, 3, 4] -> (3 x 1), [1, 2, 3, 4](1,) -> (1 x 3), [1, 2, 3, 4](2,2) -> (2 x 2)
• Frame literals: Frames can now be created using curly brackets. The syntax differs slightly for column-major and row-major format:
  • column-major: columns are given as comma-separated pairs of label : matrix, where each matrix contains all values of a column. E.g. {"col 1" : [1, 2], "col 2" : [1.1, 2.2]} (1st column of type int64_t and 2nd column of type double)
  • row-major: labels are given within brackets first, then comma-separated matrices define each row. E.g. {["col 1", "col 2"], [1, 1.1], [2, 2.2]} (equivalent to example above)

Changes:

• Matrix syntax has been extended and new syntax has been added for frames and row-major frames in DaphneDSLGrammar.g4
• New helper function buildColMatrixFromValues for handling matrix construction has been added to DaphneDSLVisitor. Furthermore, the visitor for matrices has been modified and new visitor functions for frames have been added
• castF64If has been added to ParserUtils
• 1 new instantiation for the insertRow and Fill kernels have been added to handle special cases when creating matrices of bool type
• The LanguageRef doc has been edited to reflect the above changes

Limitations:

• Row-major frame literals currently do not support labels and matrices to be passed as variables (i.e. row=[1, 2]; rowFrame={["col 1"], row};). This is because of issues in inferring the most general value type of a column which is lost if the visitor for matrix literals is called on them beforehands and casts all values to the most general value type of each row before they can be rearranged to make up the columns instead.

In the future, this could also be extended to support matrix literals where an element itself is a (flat) matrix by adding the matrix's element type to valueTypes and adding slice and rbind operations at the end of the buildColMatrixFromValues helper function.

[AlexRTer]

It appears there is an issue when running ./test.sh. Running ./test.sh ~"EwBinary*" passes all tests. However, so does ./test.sh "EwBinary*". Only when running all tests it seems to get stuck on line 35 of EwBinaryScalarTest.cpp which calls runDaphne, but I can't make out why.

[AlexRTer]

It appears there is an issue when running ./test.sh. Running ./test.sh ~"EwBinary*" passes all tests. However, so does ./test.sh "EwBinary*". Only when running all tests it seems to get stuck on line 35 of EwBinaryScalarTest.cpp which calls runDaphne, but I can't make out why.

Specifically, the runDaphne call gets stuck after calling runProgram in the same Utils.h file and reaching line 95 where it attempts to read from an stdout buffer. Nothing within the loop seems to be executed.

I still do not know why, but omitting the final ReshapeOp in the matrix literal visitor and returning colMatrix instead of its result does not cause it to get stuck and instead runs all tests successfully (except the new one specifically testing the reshaping). So the problem seems to stem from here:

mlir::Value result = static_cast<mlir::Value>(builder.create<mlir::daphne::ReshapeOp>(loc, utils.matrixOf(valueType), colMatrix, rows, cols));

[pdamme]

Thanks for this PR @AlexRTer. I can confirm the problems with the test cases, but I don't see the reason at first glance. I will look deeper into it.

[AlexRTer]

Thank you very much for your help

[AlexRTer]

I added some negative tests for matrix and frame literals.

Currently Row-major Frames allow for different column lengths, as long as the total amount of elements is a multiple of the amount of columns (the elements are simply split up evenly to all columns). For Column-major Frames, shape inference catches this case. Seeing as Column/Row -major notation is equivalent, perhaps it would be better to convert Row-major notation into Column-major notation before visiting the Frame expressions. Ideally, this would simplify Frame Literals to a single Visitor for both notations. Does Daphne currently have a way to do this?

It appears there is an issue when running ./test.sh. Running ./test.sh ~"EwBinary*" passes all tests. However, so does ./test.sh "EwBinary*". Only when running all tests it seems to get stuck on line 35 of EwBinaryScalarTest.cpp which calls runDaphne, but I can't make out why.

Thanks to c2900a3, the above issue has been resolved.

However, there seems to be an error with codegen/CodegenTest.cpp (https://github.com/daphne-eu/daphne/actions/runs/9034047874/job/24825710151#step:6:28) after the last commit that I cannot reproduce. The last commit 0e7090a only added a few matrix/frame tests and minor formatting changes. On the other hand, the tests beforehand (9bf8281) (https://github.com/daphne-eu/daphne/actions/runs/9033304615/job/24823355246) as well as my tests with the latest build both ran succesfully.

[AlexRTer]

Thanks for your feedback and editing @pdamme, I like the changes you made.

I implemented another parser rule like you suggested which fixes the issues with verifying frame literal row/column sizes, and allows expressions to be labels for frame literals of rows. Tests for matrix literals with one or no dimensions specified have also been added.

Before you merge it, I would like to refactor the code a bit and look into supporting the missing value types. Also, I still need to edit the documentation regarding support for expressions as labels.

[AlexRTer]

Matrix and Frame literals now support all value types mentioned in the documentation except for strings (i.e. si, ui with 64, 32, 8 bits, bool, f64, f32). I removed the corresponding failure test case, added tests for both matrices and frames to verify proper casting to these value types, and edited the documentation accordingly.

I added some kernel instantiations to kernels.json that are needed to support the MatrixConstant, CastScaObj, InsertRow and Reshape operations for matrices with these value types.

A remaining limitation for frame literals of rows is that rows themselves cannot be specified as expressions, unlike their elements. I am currently unaware of how to iterate over a MatrixConstant's values in the parser. This would make it possible to add support for row expressions in the frameRowMat visitor and would be preferable to creating InsertRowOps for all of its elements in my opinion.

If you think this could be left to a follow-up issue, you can merge this PR now if you would like to.