[Arith] Add internal NarrowPredicateExpression utility

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!

[Lunderberg]

Implements tvm::arith::NarrowPredicateExpression, a utility that removes free parameters from a boolean expression, such that the resulting expression being true implies that the original expression is true. For example, the predicate (0 <= i+f) && (i+f < 16), where f is a free parameter on the range 0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16).

In effect, NarrowPredicateExpression functions as a context-sentive tvm::tir::Substitute, where the value substituted is selected such that the resulting expression errs on the side of being false. This is an internal utility used as part of the simplifications for layout transformations (tracking issue link).

[Lunderberg]

Marking as a draft PR until it passes CI. This utility could also be moved to be internal to an upcoming data-flow analysis utility.

[Lunderberg]

@tvm-bot rerun

[Lunderberg]

Rebasing onto main to get the bug fix from #13067

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• Built docs for commit c6d78c1 can be found here.

_{Generated by tvm-bot}

[wrongtest-intellif]

0 <= f < 2), can be narrowed to the expression (0 <= i+0) && (i+2 < 16)

Hi~ I feel a bit confused about here: when f == 0, the i could take value in range [0, 16), thus what is the purpose to get narrowed predicate i + 2 < 16?

[Lunderberg]

Hi @wrongtest-intellif ! The goal is to produce an expression that, when true, implies that the original expression is true. For this specific example, suppose we know that i + 2 < 16. Because we also know that the free parameter f satisfies 2 < f, we know that i + 2 < i + f. Therefore, since i + f < i + 2 < 16, i+f < 16.

The NarrowPredicateExpression is part of a larger data-flow analysis that I'm preparing as part of handling padded buffer transforms. When analyzing whether a specific index corresponds to padding, conditions such as pad_left <= i+f < original_width + pad_left (identifies a non-padding data value) frequently show up. By converting these into expressions that are true regardless of the value of f, or expressions that are false regardless of the value of f, we can remove reduction axes from the overall predicate expressions.

[jwfromm]

Very cool idea and a nice implementation. Thanks @Lunderberg!