Merge remote-tracking branch 'upstream/branch-23.10' into bug/df/colu…

…mn_mismatch

.github/workflows/build.yaml

@@ -62,7 +62,7 @@ jobs:
       arch: "amd64"
       branch: ${{ inputs.branch }}
       build_type: ${{ inputs.build_type || 'branch' }}
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       date: ${{ inputs.date }}
       node_type: "gpu-v100-latest-1"
       run_script: "ci/build_docs.sh"

.github/workflows/pr.yaml

@@ -73,7 +73,7 @@ jobs:
       build_type: pull-request
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/test_java.sh"
   conda-notebook-tests:
     needs: conda-python-build
@@ -83,7 +83,7 @@ jobs:
       build_type: pull-request
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/test_notebooks.sh"
   docs-build:
     needs: conda-python-build
@@ -93,7 +93,7 @@ jobs:
       build_type: pull-request
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/build_docs.sh"
   wheel-build-cudf:
     needs: checks

.github/workflows/test.yaml

@@ -32,7 +32,7 @@ jobs:
       sha: ${{ inputs.sha }}
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/test_cpp_memcheck.sh"
   conda-python-cudf-tests:
     secrets: inherit
@@ -63,7 +63,7 @@ jobs:
       sha: ${{ inputs.sha }}
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/test_java.sh"
   conda-notebook-tests:
     secrets: inherit
@@ -75,7 +75,7 @@ jobs:
       sha: ${{ inputs.sha }}
       node_type: "gpu-v100-latest-1"
       arch: "amd64"
-      container_image: "rapidsai/ci:latest"
+      container_image: "rapidsai/ci-conda:latest"
       run_script: "ci/test_notebooks.sh"
   wheel-tests-cudf:
     secrets: inherit

conda/environments/all_cuda-118_arch-x86_64.yaml

@@ -55,8 +55,8 @@ dependencies:
 - nbsphinx
 - ninja
 - notebook
-- numba>=0.57
-- numpy>=1.21
+- numba>=0.57,<0.58
+- numpy>=1.21,<1.25
 - numpydoc
 - nvcc_linux-64=11.8
 - nvcomp==2.6.1

conda/environments/all_cuda-120_arch-x86_64.yaml

@@ -54,8 +54,8 @@ dependencies:
 - nbsphinx
 - ninja
 - notebook
-- numba>=0.57
-- numpy>=1.21
+- numba>=0.57,<0.58
+- numpy>=1.21,<1.25
 - numpydoc
 - nvcomp==2.6.1
 - nvtx>=0.2.1

conda/recipes/cudf/meta.yaml

@@ -78,8 +78,10 @@ requirements:
     - typing_extensions >=4.0.0
     - pandas >=1.3,<1.6.0dev0
     - cupy >=12.0.0
-    - numba >=0.57
-    - numpy >=1.21
+    # TODO: Pin to numba<0.58 until #14160 is resolved
+    - numba >=0.57,<0.58
+    # TODO: Pin to numpy<1.25 until cudf requires pandas 2
+    - numpy >=1.21,<1.25
     - {{ pin_compatible('pyarrow', max_pin='x.x.x') }}
     - libcudf ={{ version }}
     - {{ pin_compatible('rmm', max_pin='x.x') }}

cpp/include/cudf/rolling.hpp

@@ -199,10 +199,30 @@ struct window_bounds {
  * column of the same type as the input. Therefore it is suggested to convert integer column types
  * (especially low-precision integers) to `FLOAT32` or `FLOAT64` before doing a rolling `MEAN`.
  *
+ * Note: `preceding_window` and `following_window` could well have negative values. This yields
+ * windows where the current row might not be included at all. For instance, consider a window
+ * defined as (preceding=3, following=-1). This produces a window from 2 (i.e. 3-1) rows preceding
+ * the current row, and 1 row *preceding* the current row. For the example above, the window for
+ * row#3 is:
+ *
+ *    [ 10,  20,  10,  50,  60,  20,  30,  80,  40 ]
+ *      <--window-->   ^
+ *                     |
+ *               current_row
+ *
+ * Similarly, `preceding` could have a negative value, indicating that the window begins at a
+ * position after the current row.  It differs slightly from the semantics for `following`, because
+ * `preceding` includes the current row. Therefore:
+ *   1. preceding=1  => Window starts at the current row.
+ *   2. preceding=0  => Window starts at 1 past the current row.
+ *   3. preceding=-1 => Window starts at 2 past the current row. Etc.
+ *
  * @param[in] group_keys The (pre-sorted) grouping columns
  * @param[in] input The input column (to be aggregated)
- * @param[in] preceding_window The static rolling window size in the backward direction
- * @param[in] following_window The static rolling window size in the forward direction
+ * @param[in] preceding_window The static rolling window size in the backward direction (for
+ * positive values), or forward direction (for negative values)
+ * @param[in] following_window The static rolling window size in the forward direction (for positive
+ * values), or backward direction (for negative values)
  * @param[in] min_periods Minimum number of observations in window required to have a value,
  *                        otherwise element `i` is null.
  * @param[in] aggr The rolling window aggregation type (SUM, MAX, MIN, etc.)

cpp/src/rolling/detail/rolling.cuh

@@ -70,7 +70,22 @@ namespace cudf {
 
 namespace detail {
 
-namespace {  // anonymous
+/// Helper function to materialize preceding/following offsets.
+template <typename Calculator>
+std::unique_ptr<column> expand_to_column(Calculator const& calc,
+                                         size_type const& num_rows,
+                                         rmm::cuda_stream_view stream)
+{
+  auto window_column = cudf::make_numeric_column(
+    cudf::data_type{type_to_id<size_type>()}, num_rows, cudf::mask_state::UNALLOCATED, stream);
+
+  auto begin = cudf::detail::make_counting_transform_iterator(0, calc);
+
+  thrust::copy_n(
+    rmm::exec_policy(stream), begin, num_rows, window_column->mutable_view().data<size_type>());
+
+  return window_column;
+}
 
 /**
  * @brief Operator for applying a generic (non-specialized) rolling aggregation on a single window.
@@ -91,14 +106,14 @@ struct DeviceRolling {
 
   // operations we do support
   template <typename T = InputType, aggregation::Kind O = op>
-  DeviceRolling(size_type _min_periods, std::enable_if_t<is_supported<T, O>()>* = nullptr)
+  explicit DeviceRolling(size_type _min_periods, std::enable_if_t<is_supported<T, O>()>* = nullptr)
     : min_periods(_min_periods)
   {
   }
 
   // operations we don't support
   template <typename T = InputType, aggregation::Kind O = op>
-  DeviceRolling(size_type _min_periods, std::enable_if_t<!is_supported<T, O>()>* = nullptr)
+  explicit DeviceRolling(size_type _min_periods, std::enable_if_t<!is_supported<T, O>()>* = nullptr)
     : min_periods(_min_periods)
   {
     CUDF_FAIL("Invalid aggregation/type pair");
@@ -111,7 +126,7 @@ struct DeviceRolling {
                              mutable_column_device_view& output,
                              size_type start_index,
                              size_type end_index,
-                             size_type current_index)
+                             size_type current_index) const
   {
     using AggOp = typename corresponding_operator<op>::type;
     AggOp agg_op;
@@ -144,7 +159,7 @@ struct DeviceRolling {
 template <typename InputType, aggregation::Kind op>
 struct DeviceRollingArgMinMaxBase {
   size_type min_periods;
-  DeviceRollingArgMinMaxBase(size_type _min_periods) : min_periods(_min_periods) {}
+  explicit DeviceRollingArgMinMaxBase(size_type _min_periods) : min_periods(_min_periods) {}
 
   static constexpr bool is_supported()
   {
@@ -162,7 +177,7 @@ struct DeviceRollingArgMinMaxBase {
  */
 template <aggregation::Kind op>
 struct DeviceRollingArgMinMaxString : DeviceRollingArgMinMaxBase<cudf::string_view, op> {
-  DeviceRollingArgMinMaxString(size_type _min_periods)
+  explicit DeviceRollingArgMinMaxString(size_type _min_periods)
     : DeviceRollingArgMinMaxBase<cudf::string_view, op>(_min_periods)
   {
   }
@@ -461,8 +476,8 @@ struct agg_specific_empty_output {
   }
 };
 
-std::unique_ptr<column> empty_output_for_rolling_aggregation(column_view const& input,
-                                                             rolling_aggregation const& agg)
+static std::unique_ptr<column> empty_output_for_rolling_aggregation(column_view const& input,
+                                                                    rolling_aggregation const& agg)
 {
   // TODO:
   //  Ideally, for UDF aggregations, the returned column would match
@@ -1215,8 +1230,6 @@ struct dispatch_rolling {
   }
 };
 
-}  // namespace
-
 // Applies a user-defined rolling window function to the values in a column.
 template <typename PrecedingWindowIterator, typename FollowingWindowIterator>
 std::unique_ptr<column> rolling_window_udf(column_view const& input,

cpp/src/rolling/detail/rolling_fixed_window.cu

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022, NVIDIA CORPORATION.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -19,8 +19,9 @@
 #include <cudf/detail/aggregation/aggregation.hpp>
 #include <cudf/utilities/default_stream.hpp>
 
+#include <cudf_test/column_utilities.hpp>
+
 #include <thrust/extrema.h>
-#include <thrust/iterator/constant_iterator.h>
 
 namespace cudf::detail {
 
@@ -43,6 +44,9 @@ std::unique_ptr<column> rolling_window(column_view const& input,
   CUDF_EXPECTS((default_outputs.is_empty() || default_outputs.size() == input.size()),
                "Defaults column must be either empty or have as many rows as the input column.");
 
+  CUDF_EXPECTS(-(preceding_window - 1) <= following_window,
+               "Preceding window bounds must precede the following window bounds.");
+
   if (agg.kind == aggregation::CUDA || agg.kind == aggregation::PTX) {
     // TODO: In future, might need to clamp preceding/following to column boundaries.
     return cudf::detail::rolling_window_udf(input,
@@ -58,18 +62,22 @@ std::unique_ptr<column> rolling_window(column_view const& input,
     // Clamp preceding/following to column boundaries.
     // E.g. If preceding_window == 2, then for a column of 5 elements, preceding_window will be:
     //      [1, 2, 2, 2, 1]
-    auto const preceding_window_begin = cudf::detail::make_counting_transform_iterator(
-      0,
-      [preceding_window] __device__(size_type i) { return thrust::min(i + 1, preceding_window); });
-    auto const following_window_begin = cudf::detail::make_counting_transform_iterator(
-      0, [col_size = input.size(), following_window] __device__(size_type i) {
-        return thrust::min(col_size - i - 1, following_window);
-      });
 
+    auto const preceding_calc = [preceding_window] __device__(size_type i) {
+      return thrust::min(i + 1, preceding_window);
+    };
+
+    auto const following_calc = [col_size = input.size(),
+                                 following_window] __device__(size_type i) {
+      return thrust::min(col_size - i - 1, following_window);
+    };
+
+    auto const preceding_column = expand_to_column(preceding_calc, input.size(), stream);
+    auto const following_column = expand_to_column(following_calc, input.size(), stream);
     return cudf::detail::rolling_window(input,
                                         default_outputs,
-                                        preceding_window_begin,
-                                        following_window_begin,
+                                        preceding_column->view().begin<cudf::size_type>(),
+                                        following_column->view().begin<cudf::size_type>(),
                                         min_periods,
                                         agg,
                                         stream,