Support scalar months in add_calendrical_months, extends API to INT32 support

cpp/include/cudf/datetime.hpp

@@ -177,17 +177,52 @@ std::unique_ptr<cudf::column> day_of_year(
  *       5/30/20 00:00:01, 7/30/20 14:12:13]
  * @endcode
 
- * @param[in] timestamps cudf::column_view of timestamp type.
- * @param[in] months cudf::column_view of integer type containing the number of months to add.
+ * @throw cudf::logic_error if `timestamps` datatype is not a TIMESTAMP or if `months` datatype
+ * is not INT16 or INT32.
+ * @throw cudf::logic_error if `timestamps` column size is not equal to `months` column size.
+ *
+ * @param timestamps cudf::column_view of timestamp type.
+ * @param months cudf::column_view of integer type containing the number of months to add.
  *
  * @returns cudf::column of timestamp type containing the computed timestamps.
+ */
+std::unique_ptr<cudf::column> add_calendrical_months(
+  cudf::column_view const& timestamps,
+  cudf::column_view const& months,
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
+/**
+ * @brief  Adds or subtracts a number of months from the date time type and returns a
+ * timestamp column that is of the same type as the input `timestamps` column.
+ *
+ * For a given row, if the `timestamps` value is null, the output for that row is null.
+ * A null months scalar would result in an all null column.
+ * This method preserves the input time and the day where applicable. The date is rounded
+ * down to the last day of the month for that year, if the new day is invalid for that month.
+ *
+ * @code{.pseudo}
+ * Example:
+ * timestamps = [5/31/20 08:00:00, 6/30/20 00:00:00, 7/31/20 13:00:00]
+ * months     = -3
+ * output is [2/29/20 08:00:00, 3/30/20 00:00:00, 4/30/20 13:00:00]
+ *
+ * timestamps = [4/28/20 04:00:00, 5/30/20 01:00:00, 6/30/20 21:00:00]
+ * months     = 1
+ * output is [5/28/20 04:00:00, 6/30/20 01:00:00, 7/30/20 21:00:00]
+ * @endcode
+ *
  * @throw cudf::logic_error if `timestamps` datatype is not a TIMESTAMP or if `months` datatype
- * is not INT16.
+ * is not INT16 or INT32.
  * @throw cudf::logic_error if `timestamps` column size is not equal to `months` column size.
+ *
+ * @param timestamps cudf::column_view of timestamp type.
+ * @param months cudf::scalar of integer type containing the number of months to add.
+ *
+ * @return cudf::column of timestamp type containing the computed timestamps.
  */
 std::unique_ptr<cudf::column> add_calendrical_months(
   cudf::column_view const& timestamps,
-  cudf::column_view const& months,
+  cudf::scalar const& months,
   rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
 
 /**

cpp/include/cudf/detail/datetime.hpp

@@ -125,6 +125,18 @@ std::unique_ptr<cudf::column> add_calendrical_months(
   rmm::cuda_stream_view stream        = rmm::cuda_stream_default,
   rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
 
+/**
+ * @copydoc cudf::add_calendrical_months(cudf::column_view const&, cudf::scalar const&,
+ * rmm::mr::device_memory_resource *)
+ *
+ * @param stream CUDA stream used for device memory operations and kernel launches.
+ */
+std::unique_ptr<cudf::column> add_calendrical_months(
+  cudf::column_view const& timestamps,
+  cudf::scalar const& months,
+  rmm::cuda_stream_view stream        = rmm::cuda_stream_default,
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
 /**
  * @copydoc cudf::is_leap_year(cudf::column_view const&, rmm::mr::device_memory_resource *)
  *

cpp/src/datetime/datetime_ops.cu

@@ -20,15 +20,18 @@
 #include <cudf/column/column_view.hpp>
 #include <cudf/datetime.hpp>
 #include <cudf/detail/datetime.hpp>
+#include <cudf/detail/indexalator.cuh>
 #include <cudf/detail/null_mask.hpp>
 #include <cudf/detail/nvtx/ranges.hpp>
+#include <cudf/scalar/scalar.hpp>
 #include <cudf/table/table_view.hpp>
 #include <cudf/types.hpp>
 #include <cudf/utilities/traits.hpp>
 
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/exec_policy.hpp>
 
+#include <thrust/iterator/constant_iterator.h>
 namespace cudf {
 namespace datetime {
 namespace detail {
@@ -211,31 +214,35 @@ std::unique_ptr<column> apply_datetime_op(column_view const& column,
 }
 
 struct add_calendrical_months_functor {
-  column_view timestamp_column;
-  column_view months_column;
-  mutable_column_view output;
-
-  add_calendrical_months_functor(column_view tsc, column_view mc, mutable_column_view out)
-    : timestamp_column(tsc), months_column(mc), output(out)
-  {
-  }
-
-  template <typename Element>
-  typename std::enable_if_t<!cudf::is_timestamp_t<Element>::value, void> operator()(
-    rmm::cuda_stream_view stream) const
+  template <typename Element, typename... Args>
+  typename std::enable_if_t<!cudf::is_timestamp_t<Element>::value, std::unique_ptr<column>>
+  operator()(Args&&...) const
   {
     CUDF_FAIL("Cannot extract datetime component from non-timestamp column.");
   }
 
-  template <typename Timestamp>
-  typename std::enable_if_t<cudf::is_timestamp_t<Timestamp>::value, void> operator()(
-    rmm::cuda_stream_view stream) const
+  template <typename Timestamp, typename MonthIterator>
+  typename std::enable_if_t<cudf::is_timestamp_t<Timestamp>::value, std::unique_ptr<column>>
+  operator()(column_view timestamp_column,
+             MonthIterator months_begin,
+             rmm::cuda_stream_view stream,
+             rmm::mr::device_memory_resource* mr) const
   {
+    auto size            = timestamp_column.size();
+    auto output_col_type = timestamp_column.type();
+
+    // Return an empty column if source column is empty
+    if (size == 0) return make_empty_column(output_col_type);
+
+    auto output =
+      make_fixed_width_column(output_col_type, size, mask_state::UNALLOCATED, stream, mr);
+    auto output_mview = output->mutable_view();
+
     thrust::transform(rmm::exec_policy(stream),
                       timestamp_column.begin<Timestamp>(),
                       timestamp_column.end<Timestamp>(),
-                      months_column.begin<int16_t>(),
-                      output.begin<Timestamp>(),
+                      months_begin,
+                      output_mview.begin<Timestamp>(),
                       [] __device__(auto time_val, auto months_val) {
                         using namespace cuda::std::chrono;
                         using duration_m = duration<int32_t, months::period>;
@@ -254,6 +261,7 @@ struct add_calendrical_months_functor {
                         // Put back the time component to the date
                         return sys_days{ymd} + (time_val - days_since_epoch);
                       });
+    return output;
   }
 };
 
@@ -263,29 +271,54 @@ std::unique_ptr<column> add_calendrical_months(column_view const& timestamp_colu
                                                rmm::mr::device_memory_resource* mr)
 {
   CUDF_EXPECTS(is_timestamp(timestamp_column.type()), "Column type should be timestamp");
-  CUDF_EXPECTS(months_column.type() == data_type{type_id::INT16},
-               "Months column type should be INT16");
+  CUDF_EXPECTS(
+    months_column.type().id() == type_id::INT16 or months_column.type().id() == type_id::INT32,
+    "Months column type should be INT16 or INT32.");
   CUDF_EXPECTS(timestamp_column.size() == months_column.size(),
                "Timestamp and months column should be of the same size");
-  auto size            = timestamp_column.size();
-  auto output_col_type = timestamp_column.type();
-
-  // Return an empty column if source column is empty
-  if (size == 0) return make_empty_column(output_col_type);
-
-  auto output_col_mask =
-    cudf::detail::bitmask_and(table_view({timestamp_column, months_column}), stream, mr);
-  auto output = make_fixed_width_column(
-    output_col_type, size, std::move(output_col_mask), cudf::UNKNOWN_NULL_COUNT, stream, mr);
-
-  auto launch = add_calendrical_months_functor{
-    timestamp_column, months_column, static_cast<mutable_column_view>(*output)};
-
-  type_dispatcher(timestamp_column.type(), launch, stream);
 
+  auto const months_begin_iter =
+    cudf::detail::indexalator_factory::make_input_iterator(months_column);
+  auto output = type_dispatcher(timestamp_column.type(),
+                                add_calendrical_months_functor{},
+                                timestamp_column,
+                                months_begin_iter,
+                                stream,
+                                mr);
+
+  auto output_null_mask =
+    cudf::detail::bitmask_and(table_view{{timestamp_column, months_column}}, stream, mr);
+  output->set_null_mask(std::move(output_null_mask));
   return output;
 }
 
+std::unique_ptr<column> add_calendrical_months(column_view const& timestamp_column,
+                                               scalar const& months,
+                                               rmm::cuda_stream_view stream,
+                                               rmm::mr::device_memory_resource* mr)
+{
+  CUDF_EXPECTS(is_timestamp(timestamp_column.type()), "Column type should be timestamp");
+  CUDF_EXPECTS(months.type().id() == type_id::INT16 or months.type().id() == type_id::INT32,
+               "Months type should be INT16 or INT32");
+
+  if (months.is_valid(stream)) {
+    auto const months_begin_iter = thrust::make_permutation_iterator(
+      cudf::detail::indexalator_factory::make_input_iterator(months),
+      thrust::make_constant_iterator(0));
+    auto output = type_dispatcher(timestamp_column.type(),
+                                  add_calendrical_months_functor{},
+                                  timestamp_column,
+                                  months_begin_iter,
+                                  stream,
+                                  mr);
+    output->set_null_mask(cudf::detail::copy_bitmask(timestamp_column, stream, mr));
+    return output;
+  } else {
+    return make_timestamp_column(
+      timestamp_column.type(), timestamp_column.size(), mask_state::ALL_NULL, stream, mr);
+  }
+}
+
 std::unique_ptr<column> extract_year(column_view const& column,
                                      rmm::cuda_stream_view stream,
                                      rmm::mr::device_memory_resource* mr)
@@ -456,6 +489,14 @@ std::unique_ptr<cudf::column> add_calendrical_months(cudf::column_view const& ti
     timestamp_column, months_column, rmm::cuda_stream_default, mr);
 }
 
+std::unique_ptr<cudf::column> add_calendrical_months(cudf::column_view const& timestamp_column,
+                                                     cudf::scalar const& months,
+                                                     rmm::mr::device_memory_resource* mr)
+{
+  CUDF_FUNC_RANGE();
+  return detail::add_calendrical_months(timestamp_column, months, rmm::cuda_stream_default, mr);
+}
+
 std::unique_ptr<column> is_leap_year(column_view const& column, rmm::mr::device_memory_resource* mr)
 {
   CUDF_FUNC_RANGE();