pitrou commented on code in PR #44217:
URL: https://github.com/apache/arrow/pull/44217#discussion_r1858145672
##########
cpp/src/arrow/compute/kernels/chunked_internal.h:
##########
@@ -50,34 +56,109 @@ struct ResolvedChunk {
}
};
+// A compressed (chunk_index, index_in_chunk) pair.
+// The goal of compression is to make it fit in 64 bits, allowing in place
+// replacement of logical uint64_t indices with physical indices.
+// (see ChunkedIndexMapper)
+struct CompressedChunkLocation {
+ static constexpr int kChunkIndexBits = 24;
+ static constexpr int KIndexInChunkBits = 64 - kChunkIndexBits;
+
+ static constexpr uint64_t kMaxChunkIndex = (1ULL << kChunkIndexBits) - 1;
+ static constexpr uint64_t kMaxIndexInChunk = (1ULL << KIndexInChunkBits) - 1;
+
+ CompressedChunkLocation() = default;
+
+ constexpr uint64_t chunk_index() const { return data_ & kMaxChunkIndex; }
+ constexpr uint64_t index_in_chunk() const { return data_ >> kChunkIndexBits;
}
+
+ explicit constexpr CompressedChunkLocation(uint64_t chunk_index,
+ uint64_t index_in_chunk)
+ : data_((index_in_chunk << kChunkIndexBits) | chunk_index) {}
+
+ template <typename IndexType>
+ explicit operator TypedChunkLocation<IndexType>() {
+ return {static_cast<IndexType>(chunk_index()),
+ static_cast<IndexType>(index_in_chunk())};
+ }
+
+ private:
+ uint64_t data_;
+};
+
+static_assert(sizeof(uint64_t) == sizeof(CompressedChunkLocation));
+
class ChunkedArrayResolver {
private:
ChunkResolver resolver_;
- std::vector<const Array*> chunks_;
+ util::span<const Array* const> chunks_;
+ std::vector<const Array*> owned_chunks_;
public:
- explicit ChunkedArrayResolver(const std::vector<const Array*>& chunks)
+ explicit ChunkedArrayResolver(std::vector<const Array*>&& chunks)
+ : resolver_(chunks), chunks_(chunks), owned_chunks_(std::move(chunks)) {}
+ explicit ChunkedArrayResolver(util::span<const Array* const> chunks)
: resolver_(chunks), chunks_(chunks) {}
- ChunkedArrayResolver(ChunkedArrayResolver&& other) = default;
- ChunkedArrayResolver& operator=(ChunkedArrayResolver&& other) = default;
+ ARROW_DEFAULT_MOVE_AND_ASSIGN(ChunkedArrayResolver);
- ChunkedArrayResolver(const ChunkedArrayResolver& other) = default;
- ChunkedArrayResolver& operator=(const ChunkedArrayResolver& other) = default;
+ ChunkedArrayResolver(const ChunkedArrayResolver& other)
+ : resolver_(other.resolver_), owned_chunks_(other.owned_chunks_) {
+ // Rebind span to owned_chunks_ if necessary
+ chunks_ = owned_chunks_.empty() ? other.chunks_ : owned_chunks_;
+ }
+ ChunkedArrayResolver& operator=(const ChunkedArrayResolver& other) {
+ resolver_ = other.resolver_;
+ owned_chunks_ = other.owned_chunks_;
+ chunks_ = owned_chunks_.empty() ? other.chunks_ : owned_chunks_;
+ return *this;
+ }
ResolvedChunk Resolve(int64_t index) const {
const auto loc = resolver_.Resolve(index);
return {chunks_[loc.chunk_index], loc.index_in_chunk};
}
};
-inline std::vector<const Array*> GetArrayPointers(const ArrayVector& arrays) {
- std::vector<const Array*> pointers(arrays.size());
- std::transform(arrays.begin(), arrays.end(), pointers.begin(),
- [&](const std::shared_ptr<Array>& array) { return
array.get(); });
- return pointers;
-}
+std::vector<const Array*> GetArrayPointers(const ArrayVector& arrays);
+
+// A class that turns logical (linear) indices into physical (chunked) indices,
+// and vice-versa.
+class ChunkedIndexMapper {
+ public:
+ ChunkedIndexMapper(const std::vector<const Array*>& chunks, uint64_t*
indices_begin,
+ uint64_t* indices_end)
+ : ChunkedIndexMapper(util::span(chunks), indices_begin, indices_end) {}
+ ChunkedIndexMapper(util::span<const Array* const> chunks, uint64_t*
indices_begin,
+ uint64_t* indices_end)
+ : chunk_lengths_(GetChunkLengths(chunks)),
+ indices_begin_(indices_begin),
+ indices_end_(indices_end) {}
+ ChunkedIndexMapper(const RecordBatchVector& chunks, uint64_t* indices_begin,
+ uint64_t* indices_end)
+ : chunk_lengths_(GetChunkLengths(chunks)),
+ indices_begin_(indices_begin),
+ indices_end_(indices_end) {}
Review Comment:
Yes, on input they are chunk-partitioned. This is stated in the comment
below :)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
