-
Notifications
You must be signed in to change notification settings - Fork 902
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Use Arrow C Data Interface functions for Python interop (#15904)
This PR replaces the internals of `from_arrow` in pylibcudf with an implementation that uses the [Arrow C Data Interface](https://arrow.apache.org/docs/format/CDataInterface.html) using the [Python Capsule interface](https://arrow.apache.org/docs/format/CDataInterface/PyCapsuleInterface.html). This allows us to decouple our Python builds from using pyarrow Cython (partially, we haven't replaced the `to_arrow` conversion yet) and it will also allow us to support any other Python package that is a producer of the data interface. To support the above functionality, the following additional changes were needed in this PR: - Added the ability to produce cudf tables from `ArrowArrayStream` objects since that is what `pyarrow.Table` produces. This function is a simple wrapper around the existing `from_arrrow(ArrowArray)` API. - Added support for the large strings type, for which support has improved throughout cudf since the `from_arrow_host` API was added and for which we now require a basic overload for tests to pass. I did not add corresponding support for `from_arrow_device` to avoid ballooning the scope of this PR, so that work can be done in a follow-up. - Proper handling of `type_id::EMPTY` in concatenate because the most natural implementation of the ArrowArrayStream processing is to run `from_arrow` on each chunk and then concatenate the outputs, and from the Python side we can produce chunks of all null arrays from arrow. Contributes to #14926 Authors: - Vyas Ramasubramani (https://github.com/vyasr) Approvers: - Matthew Roeschke (https://github.com/mroeschke) - Robert Maynard (https://github.com/robertmaynard) - David Wendt (https://github.com/davidwendt) URL: #15904
- Loading branch information
Showing
14 changed files
with
466 additions
and
25 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,143 @@ | ||
| /* | ||
| * Copyright (c) 2024, NVIDIA CORPORATION. | ||
| * | ||
| * Licensed under the Apache License, Version 2.0 (the "License"); | ||
| * you may not use this file except in compliance with the License. | ||
| * You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| #include "arrow_utilities.hpp" | ||
|
|
||
| #include <cudf/column/column_factories.hpp> | ||
| #include <cudf/detail/concatenate.hpp> | ||
| #include <cudf/detail/nvtx/ranges.hpp> | ||
| #include <cudf/interop.hpp> | ||
| #include <cudf/table/table.hpp> | ||
|
|
||
| #include <rmm/cuda_stream_view.hpp> | ||
| #include <rmm/mr/device/device_memory_resource.hpp> | ||
| #include <rmm/mr/device/per_device_resource.hpp> | ||
|
|
||
| #include <nanoarrow/nanoarrow.h> | ||
| #include <nanoarrow/nanoarrow.hpp> | ||
|
|
||
| #include <memory> | ||
| #include <stdexcept> | ||
| #include <utility> | ||
| #include <vector> | ||
|
|
||
| namespace cudf { | ||
| namespace detail { | ||
|
|
||
| namespace { | ||
|
|
||
| std::unique_ptr<column> make_empty_column_from_schema(ArrowSchema const* schema, | ||
| rmm::cuda_stream_view stream, | ||
| rmm::mr::device_memory_resource* mr) | ||
| { | ||
| ArrowSchemaView schema_view; | ||
| NANOARROW_THROW_NOT_OK(ArrowSchemaViewInit(&schema_view, schema, nullptr)); | ||
|
|
||
| auto const type{arrow_to_cudf_type(&schema_view)}; | ||
| switch (type.id()) { | ||
| case type_id::EMPTY: { | ||
| return std::make_unique<column>( | ||
| data_type(type_id::EMPTY), 0, rmm::device_buffer{}, rmm::device_buffer{}, 0); | ||
| } | ||
| case type_id::LIST: { | ||
| return cudf::make_lists_column(0, | ||
| cudf::make_empty_column(data_type{type_id::INT32}), | ||
| make_empty_column_from_schema(schema->children[0], stream, mr), | ||
| 0, | ||
| {}, | ||
| stream, | ||
| mr); | ||
| } | ||
| case type_id::STRUCT: { | ||
| std::vector<std::unique_ptr<column>> child_columns; | ||
| child_columns.reserve(schema->n_children); | ||
| std::transform( | ||
| schema->children, | ||
| schema->children + schema->n_children, | ||
| std::back_inserter(child_columns), | ||
| [&](auto const& child) { return make_empty_column_from_schema(child, stream, mr); }); | ||
| return cudf::make_structs_column(0, std::move(child_columns), 0, {}, stream, mr); | ||
| } | ||
| default: { | ||
| return cudf::make_empty_column(type); | ||
| } | ||
| } | ||
| } | ||
|
|
||
| } // namespace | ||
|
|
||
| std::unique_ptr<table> from_arrow_stream(ArrowArrayStream* input, | ||
| rmm::cuda_stream_view stream, | ||
| rmm::mr::device_memory_resource* mr) | ||
| { | ||
| CUDF_EXPECTS(input != nullptr, "input ArrowArrayStream must not be NULL", std::invalid_argument); | ||
|
|
||
| // Potential future optimization: Since the from_arrow API accepts an | ||
| // ArrowSchema we're allocating one here instead of using a view, which we | ||
| // could avoid with a different underlying implementation. | ||
| ArrowSchema schema; | ||
| NANOARROW_THROW_NOT_OK(ArrowArrayStreamGetSchema(input, &schema, nullptr)); | ||
|
|
||
| std::vector<std::unique_ptr<cudf::table>> chunks; | ||
| ArrowArray chunk; | ||
| while (true) { | ||
| NANOARROW_THROW_NOT_OK(ArrowArrayStreamGetNext(input, &chunk, nullptr)); | ||
| if (chunk.release == nullptr) { break; } | ||
| chunks.push_back(from_arrow(&schema, &chunk, stream, mr)); | ||
| chunk.release(&chunk); | ||
| } | ||
| input->release(input); | ||
|
|
||
| if (chunks.empty()) { | ||
| if (schema.n_children == 0) { | ||
| schema.release(&schema); | ||
| return std::make_unique<cudf::table>(); | ||
| } | ||
|
|
||
| // If there are no chunks but the schema has children, we need to construct a suitable empty | ||
| // table. | ||
| std::vector<std::unique_ptr<cudf::column>> columns; | ||
| columns.reserve(chunks.size()); | ||
| std::transform( | ||
| schema.children, | ||
| schema.children + schema.n_children, | ||
| std::back_inserter(columns), | ||
| [&](auto const& child) { return make_empty_column_from_schema(child, stream, mr); }); | ||
| schema.release(&schema); | ||
| return std::make_unique<cudf::table>(std::move(columns)); | ||
| } | ||
|
|
||
| schema.release(&schema); | ||
|
|
||
| auto chunk_views = std::vector<table_view>{}; | ||
| chunk_views.reserve(chunks.size()); | ||
| std::transform( | ||
| chunks.begin(), chunks.end(), std::back_inserter(chunk_views), [](auto const& chunk) { | ||
| return chunk->view(); | ||
| }); | ||
| return cudf::detail::concatenate(chunk_views, stream, mr); | ||
| } | ||
|
|
||
| } // namespace detail | ||
|
|
||
| std::unique_ptr<table> from_arrow_stream(ArrowArrayStream* input, | ||
| rmm::cuda_stream_view stream, | ||
| rmm::mr::device_memory_resource* mr) | ||
| { | ||
| CUDF_FUNC_RANGE(); | ||
| return detail::from_arrow_stream(input, stream, mr); | ||
| } | ||
| } // namespace cudf |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.