Re-enable accelerated columnar-to-row path after fix in spark-rapids-jni

[thirtiseven]

Fixes #10062.

Description

PR #12699 worked around the data corruption reported in #10062 by forcing isAcceleratedTransposeSupported to false on every GPU architecture, disabling the fast AcceleratedColumnarToRowIterator path unconditionally. The root cause — an off-by-one in detail::determine_tiles that silently dropped the trailing column of the table whenever that column alone tipped the tile's row-size estimate over shmem_limit_per_tile — has since been fixed in NVIDIA/spark-rapids-jni#4493. With that JNI fix in place, the workaround is no longer needed, so this PR restores the original post-Pascal gating (Cuda.getComputeCapabilityMajor > 6).

Behaviorally this is a no-op for users — no new configuration, no plan changes. They will just see faster collect / take on wide fixed-width schemas.

Not safe to merge until NVIDIA/spark-rapids-jni#4493 has landed and the JNI SNAPSHOT that spark-rapids depends on has picked it up. Hence the [DO NOT MERGE] prefix.

Checklists

Documentation

• Updated for new or modified user-facing features or behaviors
• No user-facing change

Testing

• Added or modified tests to cover new code paths
• Covered by existing tests
  (test_hash_multiple_grpby_pivot with DATAGEN_SEED=1702610203 — the original reproducer from [BUG] test_hash_multiple_grpby_pivot DATAGEN_SEED=1702610203 fails #10062 — exercises this path on a 192-column pivot schema; the matching regression test ColumnToRowTests.PivotLikeLayout lives in spark-rapids-jni alongside the kernel fix.)
• Not required

Performance

• Tests ran and results are added in the PR description
• Issue filed with a link in the PR description
• Not required

[greptile-apps]

Greptile Summary

Removes the unconditional false workaround in isAcceleratedTransposeSupported (added in #12699 for issue #10062) and restores the original Cuda.getComputeCapabilityMajor > 6 post-Pascal gate, along with the Cuda import. This re-enables the AcceleratedColumnarToRowIterator path on Volta+ GPUs after the off-by-one in detail::determine_tiles was fixed upstream in spark-rapids-jni#4493. The PR is explicitly marked [DO NOT MERGE] pending that JNI fix landing in the consumed SNAPSHOT, which is the correct guard for this change.

Confidence Score: 5/5

Safe to merge once the upstream spark-rapids-jni#4493 fix is present in the consumed SNAPSHOT — no logic or resource management concerns in this change itself.

Single-line logic restoration with a correctly re-added import and updated copyright; only finding is a stale P2 comment. No resource lifecycle, OOM retry, or data correctness concerns introduced.

No files require special attention beyond verifying the JNI SNAPSHOT dependency.

Important Files Changed

Filename	Overview
sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuColumnarToRowExec.scala	Restores Cuda.getComputeCapabilityMajor > 6 gate in isAcceleratedTransposeSupported; adds back the Cuda import; stale workaround comment is a minor P2

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[makeIteratorFunc called] --> B{CudfRowTransitions.areAllSupported\nAND 4 < columns <= 100M?}
    B -- No --> C[ColumnarToRowIterator\n slow path]
    B -- Yes --> D{isAcceleratedTransposeSupported\nCuda.getComputeCapabilityMajor > 6?}
    D -- No\nPascal or older --> C
    D -- Yes\nVolta+ --> E[AcceleratedColumnarToRowIterator\n fast GPU path re-enabled by this PR]

Loading

_{Reviews (3): Last reviewed commit: "signoff" | Re-trigger Greptile}

[thirtiseven]

build