Add fast path for GpuNvl when lhs has no nulls

[binmahone]

• Skip isNotNull + ifElse when lhs.getNullCount == 0
• Directly return lhs.incRefCount() for better performance

This PR is broke up from #14032, please visit that to see more details

[binmahone]

build

[greptile-apps]

Greptile Summary

Optimization that adds a fast path to GpuNvl when the left-hand side column has no null values. Instead of computing isNotNull and performing conditional logic with ifElse, the code now checks if lhs.getNullCount == 0 and directly returns lhs.incRefCount(), skipping unnecessary operations.

Key changes:

• Both apply(lhs: ColumnVector, rhs: ColumnVector) and apply(lhs: ColumnVector, rhs: Scalar) overloads now include the fast path check
• When lhs.getNullCount == 0, returns lhs with incremented reference count immediately
• When nulls exist, falls back to original logic using isNotNull + ifElse
• Follows existing optimization patterns in the codebase (similar to GpuCast.scala:1417)
• Memory management is correct with proper reference counting

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The optimization is straightforward, follows existing patterns in the codebase, correctly handles memory management with incRefCount(), and preserves the original behavior when nulls are present. The fast path check is efficient and the logic is sound.
• No files require special attention

Important Files Changed

Filename	Overview
sql-plugin/src/main/scala/com/nvidia/spark/rapids/nullExpressions.scala	Added fast path optimization for GpuNvl when left-hand side has no nulls, directly returning reference-counted lhs instead of performing null check and conditional logic

Sequence Diagram

sequenceDiagram
    participant Caller
    participant GpuNvl
    participant ColumnVector as lhs ColumnVector
    participant RHS as rhs (ColumnVector/Scalar)
    
    Caller->>GpuNvl: apply(lhs, rhs)
    GpuNvl->>ColumnVector: getNullCount()
    
    alt lhs has no nulls (nullCount == 0)
        Note over GpuNvl: Fast path optimization
        GpuNvl->>ColumnVector: incRefCount()
        ColumnVector-->>GpuNvl: return lhs with incremented ref
        GpuNvl-->>Caller: return lhs
    else lhs has nulls (nullCount > 0)
        Note over GpuNvl: Original path
        GpuNvl->>ColumnVector: isNotNull()
        ColumnVector-->>GpuNvl: boolean mask
        GpuNvl->>GpuNvl: ifElse(lhs, rhs) using mask
        Note over GpuNvl: withResource closes mask
        GpuNvl-->>Caller: return result
    end

Loading

[Copilot]

Pull request overview

This PR adds a performance optimization to the GpuNvl object by introducing a fast path when the left-hand side column has no null values. Instead of performing the isNotNull check and ifElse operation, it directly returns the lhs column with an incremented reference count.

• Added fast path optimization that skips unnecessary null checks when lhs.getNullCount == 0
• Applied the same optimization to both overloaded apply methods (ColumnVector and Scalar variants)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The fast path optimization when lhs has no nulls should have explicit test coverage. Consider adding a test case in conditionals_test.py that exercises the scenario where the first argument to nvl() has nullable=False, to ensure this optimization path is tested. For example, you could add a test like test_nvl with binary_op_df using a data_gen with nullable=False.

[revans2]

The code looks good, but I don't see any performance numbers. Might be worth filing a follow on issue to look at all other null related expressions. In the past null count was something that was calculated lazily, but then CUDF changed it and we never went back and optimized things knowing that it is not cheap to use.

[revans2]

nit: If we are doing this level of optimization, would it be worth testing to see the performance improvement for using the following APIs instead.

https://github.com/rapidsai/cudf/blob/1d20c668b932cdb864389427fdaef972a5168d03/java/src/main/java/ai/rapids/cudf/ColumnView.java#L536-L548

[abellina]

Same questions as @revans2. Do note we need to update copyright.