Replace cudf::detail::valid_if with cudf::bools_to_mask

[mythrocks]

This commit is part of the continuing effort to reduce the dependency of spark-rapids-jni on cudf::detail APIs. In this commit, some of the references to cudf::detail::valid_if with cudf::bools_to_mask.

The functionality should not be altered. Existing tests ought to cover the changes.

[greptile-apps]

Greptile Summary

This PR migrates four files away from the cudf::detail::valid_if internal API to the public cudf::bools_to_mask API, continuing the effort to reduce reliance on cuDF detail headers.

• generate_input.cu: Five valid_if call sites replaced. Where valid_if could accept arbitrary iterator+predicate pairs, bools_to_mask requires a device_span<bool const>, so two call sites now materialize booleans into intermediate rmm::device_uvector<bool> buffers first (create_random_null_mask uses thrust::tabulate; the struct column path was already producing bools). Return type adapts from rmm::device_buffer to std::unique_ptr<rmm::device_buffer> using std::move(*ptr.release()).
• from_json_to_raw_map.cu: The valid_if call that applied thrust::logical_not inline is replaced by a thrust::transform into a temporary device_uvector<bool> followed by bools_to_mask.
• get_json_object.cu: Similar pattern—thrust::transform materializes validity booleans into a temporary uvector inside an IIFE, then bools_to_mask is called. Stream ordering ensures safety.
• histogram.cu: Removes an unnecessary static_cast<int8_t> when writing a bool expression to a bool array; the file already used bools_to_mask.
• One remaining valid_if usage exists in from_json_to_structs.cu, not addressed in this PR.

Confidence Score: 4/5

• This PR is a straightforward API migration with no functional changes; safe to merge after verifying tests pass.
• All replacements follow the same mechanical pattern (valid_if → bools_to_mask), the return type differences are correctly handled throughout, iterator ranges are preserved, and the changes are in benchmark/utility code with existing test coverage. The one minor formatting inconsistency (alignment in generate_input.cu) is non-functional.
• src/main/cpp/benchmarks/common/generate_input.cu has the most changes (5 call sites) and deserves the closest review.

Important Files Changed

Filename	Overview
src/main/cpp/benchmarks/common/generate_input.cu	Replaces 5 cudf::detail::valid_if call sites with cudf::bools_to_mask. Adapts return types from rmm::device_buffer to std::unique_ptr<rmm::device_buffer>, using std::move(*ptr.release()) consistently. In create_random_null_mask, materializes booleans into a uvector before calling bools_to_mask.
src/main/cpp/src/from_json_to_raw_map.cu	Replaces valid_if with an intermediate thrust::transform + bools_to_mask pattern. The logical negation that was previously inline in valid_if's predicate is now applied via thrust::transform into a temporary device_uvector<bool>.
src/main/cpp/src/get_json_object.cu	Replaces valid_if with an IIFE that materializes validity booleans via thrust::transform into a temporary device_uvector<bool>, then calls bools_to_mask. The IIFE pattern is safe because stream ordering guarantees the kernel completes before the uvector is freed.
src/main/cpp/src/histogram.cu	Minimal change: removes an unnecessary static_cast<int8_t> when assigning a bool expression to a bool element. The file already used bools_to_mask; no other functional changes.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["cudf::detail::valid_if\n(iterator + predicate → bitmask)"] -->|"Replaced by"| B["cudf::bools_to_mask\n(device_span<bool> → bitmask)"]
    
    subgraph Old["Old Pattern (detail API)"]
        A1["Iterator range + predicate functor"] --> A2["valid_if computes bools &\nproduces bitmask in one step"]
        A2 --> A3["Returns pair<rmm::device_buffer, size_type>"]
    end
    
    subgraph New["New Pattern (public API)"]
        B1["Materialize bools into\ndevice_uvector<bool>"] --> B2["Construct device_span<bool const>"]
        B2 --> B3["bools_to_mask converts\nbools → bitmask"]
        B3 --> B4["Returns pair<unique_ptr<device_buffer>, size_type>"]
        B4 --> B5["Extract via *ptr.release()"]
    end

Loading

_{Last reviewed commit: 505b3b6}

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[mythrocks]

Build

[mythrocks]

Build

[mythrocks]

c3f2550 is slightly controversial; the only way to stop using cudf::detail::valid_if is to materialize the boolean vector before packing it down to a null mask.

There might be value in requesting for a cudf::valid_if for this case, if the performance hit is too steep.

[greptile-apps]

_{4 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{4 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{4 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[mythrocks]

Build

[mythrocks]

Build

[greptile-apps]

_{4 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[ttnghia]

Please hold off a little bit. We need to discuss on mitigating the issue with code duplicates and unavoidable dependency from cudf detail namespace. We should also avoid performance impact by doing this.

[mythrocks]

Would it be viable to change this to should_be_valid, and forego the logical-not.