Skip to content

Commit

Permalink
Add cudf::explode_position (#7376)
Browse files Browse the repository at this point in the history
This PR adds support for pos_explode in cuDF. It is very similar to explode, but includes a column with the index of the array that was exploded on each row.
```
a                b
[0, 1]          100
[2]             200
[3, 4, 5]       300
```

exploded on column a would result in
```
a         pos        b
0         0         100
1         1         100
2         0         200
3         0         300
4         1         300
5         2         300
```

partially fixes #6151

Authors:
  - Mike Wilson (@hyperbolic2346)

Approvers:
  - David (@davidwendt)
  - Jake Hemstad (@jrhemstad)

URL: #7376
  • Loading branch information
hyperbolic2346 authored Feb 26, 2021
1 parent 7526be7 commit 9ae85ae
Show file tree
Hide file tree
Showing 4 changed files with 210 additions and 41 deletions.
23 changes: 22 additions & 1 deletion cpp/include/cudf/lists/lists_column_view.hpp
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2020, NVIDIA CORPORATION.
* Copyright (c) 2020-2021, NVIDIA CORPORATION.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand Down Expand Up @@ -56,6 +56,10 @@ class lists_column_view : private column_view {
using column_view::null_mask;
using column_view::offset;
using column_view::size;
using offset_type = int32_t;
static_assert(std::is_same<offset_type, size_type>::value,
"offset_type is expected to be the same as size_type.");
using offset_iterator = offset_type const*;

/**
* @brief Returns the parent column.
Expand Down Expand Up @@ -87,6 +91,23 @@ class lists_column_view : private column_view {
* @throw cudf::logic error if this is an empty column
*/
column_view get_sliced_child(rmm::cuda_stream_view stream) const;

/**
* @brief Return first offset (accounting for column offset)
*
* @return int32_t const* Pointer to the first offset
*/
offset_iterator offsets_begin() const noexcept
{
return offsets().begin<offset_type>() + offset();
}

/**
* @brief Return one past the last offset
*
* @return int32_t const* Pointer to one past the last offset
*/
offset_iterator offsets_end() const noexcept { return offsets_begin() + size(); }
};
/** @} */ // end of group
} // namespace cudf
49 changes: 46 additions & 3 deletions cpp/include/cudf/reshape.hpp
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2020, NVIDIA CORPORATION.
* Copyright (c) 2020-2021, NVIDIA CORPORATION.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand Down Expand Up @@ -125,8 +125,8 @@ std::unique_ptr<column> byte_cast(
* [null, 100],
* [15, 100],
* ```
* Note that null lists are completely removed from the output
* and nulls and empty lists inside lists are pulled out and remain.
* Note that null lists are not included in the resulting table, but nulls inside
* lists and empty lists will be represented with a null entry for that column in that row.
*
* @param input_table Table to explode.
* @param explode_column_idx Column index to explode inside the table.
Expand All @@ -139,6 +139,49 @@ std::unique_ptr<table> explode(
size_type explode_column_idx,
rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());

/**
* @brief Explodes a list column's elements and includes a position column.
*
* Any list is exploded, which means the elements of the list in each row are expanded into new rows
* in the output. The corresponding rows for other columns in the input are duplicated. A position
* column is added that has the index inside the original list for each row. Example:
* ```
* [[5,10,15], 100],
* [[20,25], 200],
* [[30], 300],
* returns
* [5, 0, 100],
* [10, 1, 100],
* [15, 2, 100],
* [20, 0, 200],
* [25, 1, 200],
* [30, 0, 300],
* ```
*
* Nulls and empty lists propagate in different ways depending on what is null or empty.
*```
* [[5,null,15], 100],
* [null, 200],
* [[], 300],
* returns
* [5, 0, 100],
* [null, 1, 100],
* [15, 2, 100],
* ```
* Note that null lists are not included in the resulting table, but nulls inside
* lists and empty lists will be represented with a null entry for that column in that row.
*
* @param input_table Table to explode.
* @param explode_column_idx Column index to explode inside the table.
* @param mr Device memory resource used to allocate the returned column's device memory.
*
* @return A new table with explode_col exploded.
*/
std::unique_ptr<table> explode_position(
table_view const& input_table,
size_type explode_column_idx,
rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());

/** @} */ // end of group

} // namespace cudf
101 changes: 66 additions & 35 deletions cpp/src/reshape/explode.cu
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@
*/

#include <cudf/column/column_device_view.cuh>
#include <cudf/detail/gather.hpp>
#include <cudf/detail/gather.cuh>
#include <cudf/detail/nvtx/ranges.hpp>
#include <cudf/lists/lists_column_view.hpp>
#include <cudf/reshape.hpp>
Expand All @@ -40,9 +40,13 @@ namespace {
* @brief Function object for exploding a column.
*/
struct explode_functor {
/**
* @brief Function object for exploding a column.
*/
template <typename T>
std::unique_ptr<table> operator()(table_view const& input_table,
size_type explode_column_idx,
size_type const explode_column_idx,
bool include_pos,
rmm::cuda_stream_view stream,
rmm::mr::device_memory_resource* mr) const
{
Expand All @@ -55,58 +59,72 @@ struct explode_functor {
template <>
std::unique_ptr<table> explode_functor::operator()<list_view>(
table_view const& input_table,
size_type explode_column_idx,
size_type const explode_column_idx,
bool include_pos,
rmm::cuda_stream_view stream,
rmm::mr::device_memory_resource* mr) const
{
/* we explode by building a gather map that includes the number of entries in each list inside
the column for each index. Interestingly, this can be done with lower_bound across the offsets
as values between the offsets will all map down to the index below. We have some off-by-one
manipulations we need to do with the output, but it's almost our gather map by itself. Once we
build the gather map we need to remove the explode column from the table and run gather on it.
Next we build the explode column, which turns out is simply lifting the child column out of the
explode column. This unrolls the top level of lists. Then we need to insert the explode column
back into the table and return it. */
lists_column_view lc{input_table.column(explode_column_idx)};
auto sliced_child = lc.get_sliced_child(stream);
rmm::device_uvector<size_type> gather_map_indices(sliced_child.size(), stream, mr);
rmm::device_uvector<size_type> gather_map_indices(sliced_child.size(), stream);

// sliced columns can make this a little tricky. We have to start iterating at the start of the
// offsets for this column, which could be > 0. Then we also have to handle rebasing the offsets
// as we go.
auto offsets = lc.offsets().begin<size_type>() + lc.offset();
// Sliced columns may require rebasing of the offsets.
auto offsets = lc.offsets_begin();
// offsets + 1 here to skip the 0th offset, which removes a - 1 operation later.
auto offsets_minus_one = thrust::make_transform_iterator(
offsets, [offsets] __device__(auto i) { return (i - offsets[0]) - 1; });
offsets + 1, [offsets] __device__(auto i) { return (i - offsets[0]) - 1; });
auto counting_iter = thrust::make_counting_iterator(0);

rmm::device_uvector<size_type> pos(include_pos ? sliced_child.size() : 0, stream, mr);

// This looks like an off-by-one bug, but what is going on here is that we need to reduce each
// result from `lower_bound` by 1 to build the correct gather map. It was pointed out that
// this can be accomplished by simply skipping the first entry and using the result of
// `lower_bound` directly.
thrust::lower_bound(rmm::exec_policy(stream),
offsets_minus_one + 1,
offsets_minus_one + lc.size() + 1,
counting_iter,
counting_iter + gather_map_indices.size(),
gather_map_indices.begin());
// result from `lower_bound` by 1 to build the correct gather map. This can be accomplished by
// skipping the first entry and using the result of `lower_bound` directly.
if (include_pos) {
thrust::transform(
rmm::exec_policy(stream),
counting_iter,
counting_iter + gather_map_indices.size(),
gather_map_indices.begin(),
[position_array = pos.data(), offsets_minus_one, offsets, offset_size = lc.size()] __device__(
auto idx) -> size_type {
auto lb_idx = thrust::lower_bound(
thrust::seq, offsets_minus_one, offsets_minus_one + offset_size, idx) -
offsets_minus_one;
position_array[idx] = idx - (offsets[lb_idx] - offsets[0]);
return lb_idx;
});
} else {
thrust::lower_bound(rmm::exec_policy(stream),
offsets_minus_one,
offsets_minus_one + lc.size(),
counting_iter,
counting_iter + gather_map_indices.size(),
gather_map_indices.begin());
}

auto select_iter = thrust::make_transform_iterator(
thrust::make_counting_iterator(0),
[explode_column_idx](size_type i) { return i >= explode_column_idx ? i + 1 : i; });
std::vector<size_type> selected_columns(select_iter, select_iter + input_table.num_columns() - 1);

auto gathered_table = cudf::detail::gather(
input_table.select(selected_columns),
column_view(data_type(type_to_id<size_type>()), sliced_child.size(), gather_map_indices.data()),
cudf::out_of_bounds_policy::DONT_CHECK,
cudf::detail::negative_index_policy::ALLOWED,
stream,
mr);
auto gathered_table = cudf::detail::gather(input_table.select(selected_columns),
gather_map_indices.begin(),
gather_map_indices.end(),
cudf::out_of_bounds_policy::DONT_CHECK,
stream,
mr);

std::vector<std::unique_ptr<column>> columns = gathered_table.release()->release();

columns.insert(columns.begin() + explode_column_idx,
std::make_unique<column>(column(sliced_child, stream, mr)));
std::make_unique<column>(sliced_child, stream, mr));

if (include_pos) {
columns.insert(columns.begin() + explode_column_idx,
std::make_unique<column>(
data_type(type_to_id<size_type>()), sliced_child.size(), pos.release()));
}

return std::make_unique<table>(std::move(columns));
}
Expand All @@ -120,13 +138,15 @@ std::unique_ptr<table> explode_functor::operator()<list_view>(
*/
std::unique_ptr<table> explode(table_view const& input_table,
size_type explode_column_idx,
bool include_pos,
rmm::cuda_stream_view stream,
rmm::mr::device_memory_resource* mr)
{
return type_dispatcher(input_table.column(explode_column_idx).type(),
explode_functor{},
input_table,
explode_column_idx,
include_pos,
stream,
mr);
}
Expand All @@ -141,7 +161,18 @@ std::unique_ptr<table> explode(table_view const& input_table,
rmm::mr::device_memory_resource* mr)
{
CUDF_FUNC_RANGE();
return detail::explode(input_table, explode_column_idx, rmm::cuda_stream_default, mr);
return detail::explode(input_table, explode_column_idx, false, rmm::cuda_stream_default, mr);
}

/**
* @copydoc cudf::explode_position(input_table,explode_column_idx,rmm::mr::device_memory_resource)
*/
std::unique_ptr<table> explode_position(table_view const& input_table,
size_type explode_column_idx,
rmm::mr::device_memory_resource* mr)
{
CUDF_FUNC_RANGE();
return detail::explode(input_table, explode_column_idx, true, rmm::cuda_stream_default, mr);
}

} // namespace cudf
Loading

0 comments on commit 9ae85ae

Please sign in to comment.