Update sort groupby to use non-atomic operation

cpp/benchmarks/groupby/group_sum_benchmark.cu

@@ -83,9 +83,11 @@ void BM_pre_sorted_sum(benchmark::State& state)
 
   auto data_it = cudf::detail::make_counting_transform_iterator(
     0, [=](cudf::size_type row) { return random_int(0, 100); });
+  auto valid_it = cudf::detail::make_counting_transform_iterator(
+    0, [=](cudf::size_type row) { return random_int(0, 100) < 90; });
 
   wrapper keys(data_it, data_it + column_size);
-  wrapper vals(data_it, data_it + column_size);
+  wrapper vals(data_it, data_it + column_size, valid_it);
 
   auto keys_table  = cudf::table_view({keys});
   auto sort_order  = cudf::sorted_order(keys_table);
@@ -111,4 +113,6 @@ BENCHMARK_DEFINE_F(Groupby, PreSorted)(::benchmark::State& state) { BM_pre_sorte
 BENCHMARK_REGISTER_F(Groupby, PreSorted)
   ->UseManualTime()
   ->Unit(benchmark::kMillisecond)
-  ->Arg(10000000);
+  ->Arg(1000000)
+  ->Arg(10000000)
+  ->Arg(100000000);

cpp/include/cudf/utilities/output_writer_iterator.cuh

@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) 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.
+ * 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.
+ */
+
+// Output writer iterator
+#pragma once
+
+#include <thrust/iterator/iterator_adaptor.h>
+
+namespace thrust {
+namespace detail {
+
+// Proxy reference that calls BinaryFunction with Iterator value and the rhs of assignment operator
+template <typename BinaryFunction, typename Iterator>
+class output_writer_iterator_proxy {
+ public:
+  __host__ __device__ output_writer_iterator_proxy(const Iterator& index_iter, BinaryFunction fun)
+    : index_iter(index_iter), fun(fun)
+  {
+  }
+  template <typename T>
+  __host__ __device__ output_writer_iterator_proxy operator=(const T& x)
+  {
+    fun(*index_iter, x);
+    return *this;
+  }
+
+ private:
+  Iterator index_iter;
+  BinaryFunction fun;
+};
+
+// Register output_writer_iterator_proxy with 'is_proxy_reference' from
+// type_traits to enable its use with algorithms.
+template <class BinaryFunction, class Iterator>
+struct is_proxy_reference<output_writer_iterator_proxy<BinaryFunction, Iterator>>
+  : public thrust::detail::true_type {
+};
+
+}  // namespace detail
+}  // namespace thrust
+
+namespace cudf {
+/**
+ * @brief Transform output iterator with custom writer binary function which takes index and value.
+ *
+ * @code {.cpp}
+ * #include <cudf/utilities/output_writer_iterator.cuh>
+ * #include <thrust/device_vector.h>
+ * #include <thrust/iterator/counting_iterator.h>
+ * #include <thrust/iterator/transform_iterator.h>
+ *
+ * struct set_bits_field {
+ *   int* bitfield;
+ *   __device__ inline void set_bit(size_t bit_index)
+ *   {
+ *     atomicOr(&bitfield[bit_index/32], (int{1} << (bit_index % 32)));
+ *   }
+ *   __device__ inline void clear_bit(size_t bit_index)
+ *   {
+ *     atomicAnd(&bitfield[bit_index / 32], ~(int{1} << (bit_index % 32)));
+ *   }
+ *   // Index, value
+ *   __device__ void operator()(size_t i, bool x)
+ *   {
+ *     if (x)
+ *       set_bit(i);
+ *     else
+ *       clear_bit(i);
+ *   }
+ * };
+ *
+ * thrust::device_vector<int> v(1, 0x0000);
+ * auto result_begin = cudf::make_output_writer_iterator(thrust::make_counting_iterator(0),
+ *                                                 set_bits_field{v.data().get()});
+ * auto value = thrust::make_transform_iterator(thrust::make_counting_iterator(0), [] __device__
+ * (int x) { return x%2;
+ * });
+ * thrust::copy(thrust::device, value, value+32, result_begin);
+ *
+ * #include <cudf/utilities/output_writer_iterator.cuh>
+ * #include <thrust/device_vector.h>
+ * #include <thrust/iterator/counting_iterator.h>
+ * #include <thrust/iterator/transform_iterator.h>
+ *
+ * struct set_bits_field {
+ *   int* bitfield;
+ *   __device__ inline void set_bit(size_t bit_index)
+ *   {
+ *     atomicOr(&bitfield[bit_index/32], (int{1} << (bit_index % 32)));
+ *   }
+ *   __device__ inline void clear_bit(size_t bit_index)
+ *   {
+ *     atomicAnd(&bitfield[bit_index / 32], ~(int{1} << (bit_index % 32)));
+ *   }
+ *   // Index, value
+ *   __device__ void operator()(size_t i, bool x)
+ *   {
+ *     if (x)
+ *       set_bit(i);
+ *     else
+ *       clear_bit(i);
+ *   }
+ * };
+ *
+ * thrust::device_vector<int> v(1, 0x0000);
+ * auto result_begin = cudf::make_output_writer_iterator(thrust::make_counting_iterator(0),
+ *                                                 set_bits_field{v.data().get()});
+ * auto value = thrust::make_transform_iterator(thrust::make_counting_iterator(0),
+ *   [] __device__ (int x) {   return x%2; });
+ * thrust::copy(thrust::device, value, value+32, result_begin);
+ * int(v[0]); // returns 0xaaaaaaaa;
+ * @endcode
+ *
+ *
+ * @tparam BinaryFunction Binary function to be called with the Iterator value and the rhs of
+ * assignment operator.
+ * @tparam Iterator iterator type that acts as index of the output.
+ */
+template <typename BinaryFunction, typename Iterator>
+class output_writer_iterator
+  : public thrust::iterator_adaptor<
+      output_writer_iterator<BinaryFunction, Iterator>,
+      Iterator,
+      thrust::use_default,
+      thrust::use_default,
+      thrust::use_default,
+      thrust::detail::output_writer_iterator_proxy<BinaryFunction, Iterator>> {
+ public:
+  // parent class.
+  typedef thrust::iterator_adaptor<
+    output_writer_iterator<BinaryFunction, Iterator>,
+    Iterator,
+    thrust::use_default,
+    thrust::use_default,
+    thrust::use_default,
+    thrust::detail::output_writer_iterator_proxy<BinaryFunction, Iterator>>
+    super_t;
+  // friend thrust::iterator_core_access to allow it access to the private interface dereference()
+  friend class thrust::iterator_core_access;
+  __host__ __device__ output_writer_iterator(Iterator const& x, BinaryFunction fun)
+    : super_t(x), fun(fun)
+  {
+  }
+
+ private:
+  BinaryFunction fun;
+
+  // thrust::iterator_core_access accesses this function
+  __host__ __device__ typename super_t::reference dereference() const
+  {
+    return thrust::detail::output_writer_iterator_proxy<BinaryFunction, Iterator>(
+      this->base_reference(), fun);
+  }
+};
+
+template <typename BinaryFunction, typename Iterator>
+output_writer_iterator<BinaryFunction, Iterator> __host__ __device__
+make_output_writer_iterator(Iterator out, BinaryFunction fun)
+{
+  return output_writer_iterator<BinaryFunction, Iterator>(out, fun);
+}  // end make_output_writer_iterator
+}  // namespace cudf

cpp/src/groupby/sort/group_argmax.cu

@@ -33,7 +33,10 @@ std::unique_ptr<column> group_argmax(column_view const& values,
                                      rmm::cuda_stream_view stream,
                                      rmm::mr::device_memory_resource* mr)
 {
-  auto indices = type_dispatcher(values.type(),
+  auto values_type = cudf::is_dictionary(values.type())
+                       ? dictionary_column_view(values).keys().type()
+                       : values.type();
+  auto indices     = type_dispatcher(values_type,
                                  reduce_functor<aggregation::ARGMAX>{},
                                  values,
                                  num_groups,

cpp/src/groupby/sort/group_argmin.cu

@@ -33,7 +33,10 @@ std::unique_ptr<column> group_argmin(column_view const& values,
                                      rmm::cuda_stream_view stream,
                                      rmm::mr::device_memory_resource* mr)
 {
-  auto indices = type_dispatcher(values.type(),
+  auto values_type = cudf::is_dictionary(values.type())
+                       ? dictionary_column_view(values).keys().type()
+                       : values.type();
+  auto indices     = type_dispatcher(values_type,
                                  reduce_functor<aggregation::ARGMIN>{},
                                  values,
                                  num_groups,

cpp/src/groupby/sort/group_max.cu

@@ -27,13 +27,11 @@ std::unique_ptr<column> group_max(column_view const& values,
                                   rmm::cuda_stream_view stream,
                                   rmm::mr::device_memory_resource* mr)
 {
-  return type_dispatcher(values.type(),
-                         reduce_functor<aggregation::MAX>{},
-                         values,
-                         num_groups,
-                         group_labels,
-                         stream,
-                         mr);
+  auto values_type = cudf::is_dictionary(values.type())
+                       ? dictionary_column_view(values).keys().type()
+                       : values.type();
+  return type_dispatcher(
+    values_type, reduce_functor<aggregation::MAX>{}, values, num_groups, group_labels, stream, mr);
 }
 
 }  // namespace detail

cpp/src/groupby/sort/group_min.cu

@@ -27,13 +27,11 @@ std::unique_ptr<column> group_min(column_view const& values,
                                   rmm::cuda_stream_view stream,
                                   rmm::mr::device_memory_resource* mr)
 {
-  return type_dispatcher(values.type(),
-                         reduce_functor<aggregation::MIN>{},
-                         values,
-                         num_groups,
-                         group_labels,
-                         stream,
-                         mr);
+  auto values_type = cudf::is_dictionary(values.type())
+                       ? dictionary_column_view(values).keys().type()
+                       : values.type();
+  return type_dispatcher(
+    values_type, reduce_functor<aggregation::MIN>{}, values, num_groups, group_labels, stream, mr);
 }
 
 }  // namespace detail