[GitHub] [arrow-rs] alamb commented on a diff in pull request #1499: Alternative implementation of nullif kernel by slicing nested buffers

[GitBox <gi...@apache.org>]

alamb commented on code in PR #1499:
URL: https://github.com/apache/arrow-rs/pull/1499#discussion_r848538708


##########
arrow/src/compute/kernels/boolean.rs:
##########
@@ -575,10 +577,242 @@ where
     Ok(PrimitiveArray::<T>::from(data))
 }
 
+/// Creates a (mostly) zero-copy slice of the given buffers so that they can be combined
+/// in the same array with other buffers that start at offset 0.
+/// The only buffers that need an actual copy are booleans (if they are not byte-aligned)
+/// and list/binary/string offsets because the arrow implementation requires them to start at 0.
+/// This is useful when a kernel calculates a new validity bitmap but wants to reuse other buffers.
+fn slice_buffers(
+    buffers: &[Buffer],
+    offset: usize,
+    len: usize,
+    data_type: &DataType,
+    child_data: &[ArrayData],
+) -> (Vec<Buffer>, Vec<ArrayData>) {
+    use std::mem::size_of;
+
+    if offset == 0 {
+        return (buffers.to_vec(), child_data.to_vec());
+    }
+
+    // we only need to do something special to child data in 2 of the match branches
+    let mut result_child_data = None;
+
+    let result_buffers = match data_type {
+        DataType::Boolean => vec![buffers[0].bit_slice(offset, len)],
+        DataType::Int8 | DataType::UInt8 => {
+            vec![buffers[0].slice(offset * size_of::<u8>())]
+        }
+        DataType::Int16 | DataType::UInt16 | DataType::Float16 => {
+            vec![buffers[0].slice(offset * size_of::<u16>())]
+        }
+        DataType::Int32 | DataType::UInt32 | DataType::Float32 => {
+            vec![buffers[0].slice(offset * size_of::<u32>())]
+        }
+        DataType::Int64 | DataType::UInt64 | DataType::Float64 => {
+            vec![buffers[0].slice(offset * size_of::<u64>())]
+        }
+        DataType::Timestamp(_, _) | DataType::Duration(_) => {
+            vec![buffers[0].slice(offset * size_of::<u64>())]
+        }
+        DataType::Date32 | DataType::Time32(_) => {
+            vec![buffers[0].slice(offset * size_of::<u32>())]
+        }
+        DataType::Date64 | DataType::Time64(_) => {
+            vec![buffers[0].slice(offset * size_of::<u64>())]
+        }
+        DataType::Interval(IntervalUnit::YearMonth) => {
+            vec![buffers[0].slice(offset * size_of::<u32>())]
+        }
+        DataType::Interval(IntervalUnit::DayTime) => {
+            vec![buffers[0].slice(offset * 2 * size_of::<u32>())]
+        }
+        DataType::Interval(IntervalUnit::MonthDayNano) => {
+            vec![buffers[0].slice(offset * (2 * size_of::<u32>() + size_of::<u64>()))]
+        }
+        DataType::Decimal(_, _) => vec![buffers[0].slice(offset * size_of::<i128>())],
+        DataType::Dictionary(key_type, _) => match key_type.as_ref() {
+            DataType::Int8 | DataType::UInt8 => {
+                vec![buffers[0].slice(offset * size_of::<u8>())]
+            }
+            DataType::Int16 | DataType::UInt16 => {
+                vec![buffers[0].slice(offset * size_of::<u16>())]
+            }
+            DataType::Int32 | DataType::UInt32 => {
+                vec![buffers[0].slice(offset * size_of::<u32>())]
+            }
+            DataType::Int64 | DataType::UInt64 => {
+                vec![buffers[0].slice(offset * size_of::<u64>())]
+            }
+            _ => unreachable!(),
+        },
+        DataType::List(_) => {
+            // safe because for List the first buffer is guaranteed to contain i32 offsets
+            let offsets = unsafe { &buffers[0].typed_data()[offset..] };
+            let first_offset = offsets[0] as usize;
+            let last_offset = offsets[len] as usize;
+            let nested_len = last_offset - first_offset;
+
+            // since we calculate a new offset buffer starting from 0 we also have to slice the child data
+            result_child_data = Some(
+                child_data
+                    .iter()
+                    .map(|d| d.slice(first_offset, nested_len))
+                    .collect(),
+            );
+            vec![offset_buffer_slice::<i32>(offsets, len)]
+        }
+        DataType::LargeList(_) => {
+            // safe because for LargeList the first buffer is guaranteed to contain i64 offsets
+            let offsets = unsafe { &buffers[0].typed_data()[offset..] };
+            let first_offset = offsets[0] as usize;
+            let last_offset = offsets[len] as usize;
+            let nested_len = last_offset - first_offset;
+            // since we calculate a new offset buffer starting from 0 we also have to slice the child data
+
+            result_child_data = Some(
+                child_data
+                    .iter()
+                    .map(|d| d.slice(first_offset, nested_len))
+                    .collect(),
+            );
+            vec![offset_buffer_slice::<i64>(offsets, len)]
+        }
+        DataType::Binary | DataType::Utf8 => {
+            // safe because for Binary/Utf8 the first buffer is guaranteed to contain i32 offsets
+            let offsets = unsafe { &buffers[0].typed_data()[offset..] };
+            let first_offset = offsets[0] as usize;
+            vec![
+                offset_buffer_slice::<i32>(offsets, len),
+                buffers[1].slice(first_offset * size_of::<i32>()),
+            ]
+        }
+        DataType::LargeBinary | DataType::LargeUtf8 => {
+            // safe because for LargeBinary/LargeUtf8 the first buffer is guaranteed to contain i64 offsets
+            let offsets = unsafe { &buffers[0].typed_data()[offset..] };
+            let first_offset = offsets[0] as usize;
+            vec![
+                offset_buffer_slice::<i64>(offsets, len),
+                buffers[1].slice(first_offset * size_of::<i64>()),
+            ]
+        }
+        DataType::FixedSizeBinary(size) => {
+            vec![buffers[0].slice(offset * (*size as usize))]
+        }
+        DataType::FixedSizeList(_, _size) => {
+            // TODO: should this actually slice the child arrays?
+            vec![]
+        }
+        DataType::Struct(_) => {
+            // TODO: should this actually slice the child arrays?
+            vec![]
+        }
+        DataType::Union(..) => {
+            // TODO: verify this is actually correct
+            if buffers.len() > 1 {
+                // dense union, type ids and offsets
+                vec![
+                    buffers[0].slice(offset * size_of::<i8>()),
+                    buffers[1].slice(offset * size_of::<u32>()),
+                ]
+            } else {
+                // sparse union with only type ids
+                // should this also slice the child arrays?
+                vec![buffers[0].slice(offset * size_of::<i8>())]
+            }
+        }
+        DataType::Map(_, _) => {
+            // TODO: verify this is actually correct
+            result_child_data =
+                Some(child_data.iter().map(|d| d.slice(offset, len)).collect());
+            vec![]
+        }
+        DataType::Null => vec![],
+    };
+
+    (
+        result_buffers,
+        result_child_data.unwrap_or_else(|| child_data.to_vec()),
+    )
+}
+
+// ListArray::from(ArrayData) expected offsets to always start at 0 so we can't simply make a slice of the buffer,
+// but instead have to calculate a new buffer
+fn offset_buffer_slice<T: OffsetSizeTrait>(
+    original_offsets: &[T],
+    array_len: usize,
+) -> Buffer {
+    // offset buffer has 1 element more than the corresponding data buffer pointing yet another offset for the end of the buffer
+    let len_in_bytes = (array_len + 1) * std::mem::size_of::<T>();
+    let mut offset_buffer =
+        MutableBuffer::new(len_in_bytes).with_bitset(len_in_bytes, false);
+    let offset_slice = unsafe { offset_buffer.typed_data_mut::<T>() };
+    let offset_start = original_offsets[0];
+    offset_slice
+        .iter_mut()
+        .zip(original_offsets.iter())
+        .for_each(|(output, input)| {
+            *output = *input - offset_start;
+        });
+
+    offset_buffer.into()
+}
+
+pub fn nullif_alternative(
+    array: &ArrayRef,
+    condition: &BooleanArray,
+) -> Result<ArrayRef> {
+    if array.len() != condition.len() {
+        return Err(ArrowError::ComputeError(
+            "Inputs to conditional kernels must have the same length".to_string(),
+        ));
+    }
+
+    let condition_buffer = combine_option_buffers(
+        condition.data().buffers().first(),
+        condition.offset(),
+        condition.data().null_buffer(),
+        condition.offset(),
+        condition.len(),
+    )
+    .map(|b| !&b);
+
+    let result_valid_buffer = combine_option_buffers(
+        condition_buffer.as_ref(),
+        0,
+        array.data().null_buffer(),
+        array.offset(),
+        condition.len(),
+    );
+
+    let (result_data_buffers, result_child_data) = slice_buffers(

Review Comment:
   Perhaps `slice()` belongs on `ArrayData` (which has both the buffers and datatypes)



-- 
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: github-unsubscribe@arrow.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org