[FEA] Add support for scalar column views in AST expressions and JIT execution by lamarrr · Pull Request #23615 · NVIDIA/cudf

lamarrr · 2026-08-10T23:29:01Z

Description

This PR allows AST literals used by the JIT executor to reference a preconstructed
scalar_column_view. Previously, every scalar literal was materialized as a one-row column on
every JIT transform call. That setup cost dominates workloads that repeatedly evaluate expressions
over wide tables or small row batches.

The change:

Extends cudf::ast::literal to hold either an existing cudf::scalar or a non-owning
scalar_column_view.
Updates AST literal type and validity handling for both representations.
Updates the row-IR converter to pass prepared scalar-column views directly to cudf::transform
while retaining ownership of columns materialized for the existing scalar path.
Adds ast_jit_wide_table to TRANSFORM_NVBENCH, covering:
- table width;
- rows per transform call;
- total rows;
- expression depth;
- ast, jit, and jit-opt executors.

In the benchmark, jit uses AST literals backed by cudf::scalar, while jit-opt uses
equivalent literals backed by preconstructed scalar_column_view objects.

A literal constructed from a scalar_column_view is non-owning. The referenced one-row column must
remain alive for the lifetime of the literal and its evaluation.

Performance

Measured on an NVIDIA RTX A6000 with 262,144 total rows and a table width of one. Expression depth
and rows per transform call were swept from 1 to 128 and 512 to 131,072, respectively.

Expression depth	`jit-opt` speedup over `jit`
1	1.79-1.95x
2	2.90-2.95x
4	4.82-4.98x
8	5.83-7.08x
16	11.24-12.01x
32	17.04-19.71x
64	19.74-22.66x
128	24.28-27.24x

At expression depth 128:

512 rows per call improved from 2,363.69 ms to 86.76 ms (27.24x).
131,072 rows per call improved from 9.30 ms to 0.383 ms (24.28x).

A matched Nsight Systems profile at depth 16 and 1,024 rows per call showed that both paths execute
the same 256 JIT transform kernels with effectively identical transform-kernel time (0.601 ms versus
0.599 ms). Reusing scalar-column views removed 4,096 scalar copy kernels and 4,096 async copies,
reducing the profiled benchmark range from 189.9 ms to 21.1 ms (8.98x).

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.

- Introduced `scalar_column_view` handling in the `literal` class. - Updated `row_ir` to accommodate both scalar columns and scalar column views. - Enhanced input handling in JIT execution to support new scalar types. - Added a new benchmark for wide table transformations in NVBench.

coderabbitai · 2026-08-10T23:32:33Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 88c873a5-f890-48b0-afdb-cde919d47d05

📥 Commits

Reviewing files that changed from the base of the PR and between 9c734e4 and 2d2085d.

📒 Files selected for processing (1)

cpp/benchmarks/transform/transform_widetable.cpp

🚧 Files skipped from review as they are similar to previous changes (1)

cpp/benchmarks/transform/transform_widetable.cpp

📝 Walkthrough

Summary by CodeRabbit

New Features
- Added support for scalar column views in expression evaluation and JIT execution.
- Expanded transform benchmarking to cover wide tables, configurable batch sizes, row counts, expression depth, and execution modes.
- Added support for nested addition expression benchmarks using scalar and scalar-column inputs.

Walkthrough

The PR adds scalar column view support to AST literals and Row IR inputs. It also adds and registers a configurable wide-table transform benchmark for AST, JIT, and optimized JIT executors.

Changes

AST and JIT transform support

Layer / File(s)	Summary
Literal scalar representation `cpp/include/cudf/ast/expressions.hpp`	`literal` now stores either scalar metadata or a `scalar_column_view`. Accessors and validity checks dispatch between both representations.
Row IR scalar input integration `cpp/src/jit/row_ir.hpp`, `cpp/src/jit/row_ir.cpp`	Row IR inputs now accept owned scalar columns and scalar column views. Input conversion, nullability, column extraction, and transform input assembly handle both variants.
Wide-table transform benchmark `cpp/benchmarks/transform/transform_widetable.cpp`, `cpp/benchmarks/CMakeLists.txt`	Adds the configurable `ast_jit_wide_table` benchmark and includes it in the NVBench transform target.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Suggested labels: feature request

Suggested reviewers: mattgara, misiugodfrey

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 34.62% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely summarizes the main change: adding scalar column view support to AST expressions and JIT execution.
Description check	✅ Passed	The description directly explains the scalar column view changes, JIT integration, lifetime requirement, benchmark additions, and performance results.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

Create stacked PR
Commit on current branch

🧪 Generate unit tests (beta)

Create PR with unit tests

_{Comment @coderabbitai help to get the list of available commands.}

coderabbitai

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cpp/benchmarks/transform/transform_widetable.cpp`:
- Around line 75-76: Run clang-format on the create_sequence_table call in the
benchmark setup, ensuring the wrapped arguments align according to the project's
formatting rules. Limit the change to formatting in transform_widetable.cpp.
- Around line 21-25: Add the direct standard-library headers <utility> for
std::move and <cstddef> for std::size_t in the include section of
transform_widetable.cpp, alongside the existing headers.

In `@cpp/include/cudf/ast/expressions.hpp`:
- Around line 303-308: Update the Doxygen documentation for the
scalar_column_view constructor of literal to explicitly note that the referenced
one-row column must outlive the literal and every JIT evaluation using it,
reflecting its non-owning lifetime requirement.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 9d6a2d79-4f72-48f3-81ef-fc3b9708f791

📥 Commits

Reviewing files that changed from the base of the PR and between baea696 and 9c734e4.

📒 Files selected for processing (5)

cpp/benchmarks/CMakeLists.txt
cpp/benchmarks/transform/transform_widetable.cpp
cpp/include/cudf/ast/expressions.hpp
cpp/src/jit/row_ir.cpp
cpp/src/jit/row_ir.hpp

coderabbitai · 2026-08-10T23:32:36Z

+  /**
+   * @brief Construct a new literal object.
+   *
+   * @param value A scalar column view value
+   */
+  literal(scalar_column_view value) : scalar{std::move(value)} {}


🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

Document the non-owning lifetime requirement.

literal retains a non-owning scalar_column_view. cpp/src/jit/row_ir.cpp later forwards that view into transform_args.inputs. If the referenced column is destroyed before JIT evaluation, the transform can access invalid device memory.

Add a Doxygen note that the referenced one-row column must outlive the literal and every JIT evaluation that uses it.

Proposed documentation change

/** * `@brief` Construct a new literal object. * * `@param` value A scalar column view value + * `@note` `value` is non-owning. The referenced one-row column must outlive this + * `literal` and every evaluation that uses it. */

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

/**

* @brief Construct a new literal object.

*

* @param value A scalar column view value

*/

literal(scalar_column_view value) : scalar{std::move(value)} {}

/**

* `@brief` Construct a new literal object.

*

* `@param` value A scalar column view value

* `@note` `value` is non-owning. The referenced one-row column must outlive this

* `literal` and every evaluation that uses it.

*/

literal(scalar_column_view value) : scalar{std::move(value)} {}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@cpp/include/cudf/ast/expressions.hpp` around lines 303 - 308, Update the Doxygen documentation for the scalar_column_view constructor of literal to explicitly note that the referenced one-row column must outlive the literal and every JIT evaluation using it, reflecting its non-owning lifetime requirement.

Source: Coding guidelines

@lamarrr Can you confirm this? I don't recall the ownership model of scalar_column_view but I think this is incorrect.

That's the lifetime requirement for all cuDF column/scalar view types. It would be noise to re-document that at every usage site

lamarrr requested review from a team as code owners August 10, 2026 23:29

lamarrr requested review from mattgara and misiugodfrey August 10, 2026 23:29

github-actions Bot assigned lamarrr Aug 10, 2026

github-actions Bot added libcudf Affects libcudf (C++/CUDA) code. CMake CMake build issue labels Aug 10, 2026

lamarrr changed the title ~~Add support for scalar column views in AST expressions and JIT execution~~ [FEA] Add support for scalar column views in AST expressions and JIT execution Aug 10, 2026

lamarrr removed libcudf Affects libcudf (C++/CUDA) code. CMake CMake build issue labels Aug 10, 2026

coderabbitai Bot reviewed Aug 10, 2026

View reviewed changes

Refactor formatting and include directives in transform_widetable.cpp

2d2085d

github-actions Bot added libcudf Affects libcudf (C++/CUDA) code. CMake CMake build issue labels Aug 10, 2026

lamarrr added this to the Wide-Table JIT Performance Improvement milestone Aug 11, 2026

lamarrr added feature request New feature or request non-breaking Non-breaking change labels Aug 11, 2026

Navigation Menu

Uh oh!

[FEA] Add support for scalar column views in AST expressions and JIT execution - #23615

[FEA] Add support for scalar column views in AST expressions and JIT execution#23615
lamarrr wants to merge 2 commits into
NVIDIA:mainfrom
lamarrr:literal-opt

lamarrr commented Aug 10, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Aug 10, 2026 •

edited

Loading

Summary by CodeRabbit

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

Uh oh!

coderabbitai Bot Aug 10, 2026

Uh oh!

bdice Aug 11, 2026 •

edited

Loading

Uh oh!

lamarrr Aug 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

lamarrr commented Aug 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Performance

Checklist

Uh oh!

coderabbitai Bot commented Aug 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

coderabbitai Bot Aug 10, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Aug 11, 2026 • 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!

lamarrr Aug 11, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

lamarrr commented Aug 10, 2026 •

edited

Loading

coderabbitai Bot commented Aug 10, 2026 •

edited

Loading

bdice Aug 11, 2026 •

edited

Loading