/home/runner/work/cccl/cccl/cub/cub/device/device_reduce.cuh
File members: /home/runner/work/cccl/cccl/cub/cub/device/device_reduce.cuh
/******************************************************************************
* Copyright (c) 2011, Duane Merrill. All rights reserved.
* Copyright (c) 2011-2022, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the
* names of its contributors may be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
* DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
******************************************************************************/
#pragma once
#include <cub/config.cuh>
#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
# pragma GCC system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
# pragma clang system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
# pragma system_header
#endif // no system header
#include <cub/detail/choose_offset.cuh>
#include <cub/detail/nvtx.cuh>
#include <cub/device/dispatch/dispatch_reduce.cuh>
#include <cub/device/dispatch/dispatch_reduce_by_key.cuh>
#include <cub/iterator/arg_index_input_iterator.cuh>
#include <cub/util_deprecated.cuh>
#include <iterator>
#include <limits>
CUB_NAMESPACE_BEGIN
struct DeviceReduce
{
template <typename InputIteratorT, typename OutputIteratorT, typename ReductionOpT, typename T, typename NumItemsT>
CUB_RUNTIME_FUNCTION static cudaError_t Reduce(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
NumItemsT num_items,
ReductionOpT reduction_op,
T init,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::Reduce");
// Signed integer type for global offsets
using OffsetT = detail::choose_offset_t<NumItemsT>;
return DispatchReduce<InputIteratorT, OutputIteratorT, OffsetT, ReductionOpT, T>::Dispatch(
d_temp_storage, temp_storage_bytes, d_in, d_out, static_cast<OffsetT>(num_items), reduction_op, init, stream);
}
template <typename InputIteratorT, typename OutputIteratorT, typename ReductionOpT, typename T>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t Reduce(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
ReductionOpT reduction_op,
T init,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Reduce<InputIteratorT, OutputIteratorT, ReductionOpT, T>(
d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, reduction_op, init, stream);
}
template <typename InputIteratorT, typename OutputIteratorT, typename NumItemsT>
CUB_RUNTIME_FUNCTION static cudaError_t
Sum(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
NumItemsT num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::Sum");
// Signed integer type for global offsets
using OffsetT = detail::choose_offset_t<NumItemsT>;
// The output value type
using OutputT = cub::detail::non_void_value_t<OutputIteratorT, cub::detail::value_t<InputIteratorT>>;
using InitT = OutputT;
return DispatchReduce<InputIteratorT, OutputIteratorT, OffsetT, cub::Sum, InitT>::Dispatch(
d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
static_cast<OffsetT>(num_items),
cub::Sum(),
InitT{}, // zero-initialize
stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t
Sum(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Sum<InputIteratorT, OutputIteratorT>(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream);
}
template <typename InputIteratorT, typename OutputIteratorT, typename NumItemsT>
CUB_RUNTIME_FUNCTION static cudaError_t
Min(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
NumItemsT num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::Min");
// Signed integer type for global offsets
using OffsetT = detail::choose_offset_t<NumItemsT>;
// The input value type
using InputT = cub::detail::value_t<InputIteratorT>;
using InitT = InputT;
return DispatchReduce<InputIteratorT, OutputIteratorT, OffsetT, cub::Min, InitT>::Dispatch(
d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
static_cast<OffsetT>(num_items),
cub::Min(),
// replace with
// std::numeric_limits<T>::max() when
// C++11 support is more prevalent
Traits<InitT>::Max(),
stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t
Min(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Min<InputIteratorT, OutputIteratorT>(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_RUNTIME_FUNCTION static cudaError_t ArgMin(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::ArgMin");
// Signed integer type for global offsets
using OffsetT = int;
// The input type
using InputValueT = cub::detail::value_t<InputIteratorT>;
// The output tuple type
using OutputTupleT = cub::detail::non_void_value_t<OutputIteratorT, KeyValuePair<OffsetT, InputValueT>>;
using AccumT = OutputTupleT;
using InitT = detail::reduce::empty_problem_init_t<AccumT>;
// The output value type
using OutputValueT = typename OutputTupleT::Value;
// Wrapped input iterator to produce index-value <OffsetT, InputT> tuples
using ArgIndexInputIteratorT = ArgIndexInputIterator<InputIteratorT, OffsetT, OutputValueT>;
ArgIndexInputIteratorT d_indexed_in(d_in);
// Initial value
// TODO Address https://github.com/NVIDIA/cub/issues/651
InitT initial_value{AccumT(1, Traits<InputValueT>::Max())};
return DispatchReduce<ArgIndexInputIteratorT, OutputIteratorT, OffsetT, cub::ArgMin, InitT, AccumT>::Dispatch(
d_temp_storage, temp_storage_bytes, d_indexed_in, d_out, num_items, cub::ArgMin(), initial_value, stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t ArgMin(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return ArgMin<InputIteratorT, OutputIteratorT>(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream);
}
template <typename InputIteratorT, typename OutputIteratorT, typename NumItemsT>
CUB_RUNTIME_FUNCTION static cudaError_t
Max(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
NumItemsT num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::Max");
// Signed integer type for global offsets
using OffsetT = detail::choose_offset_t<NumItemsT>;
// The input value type
using InputT = cub::detail::value_t<InputIteratorT>;
using InitT = InputT;
return DispatchReduce<InputIteratorT, OutputIteratorT, OffsetT, cub::Max, InitT>::Dispatch(
d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
static_cast<OffsetT>(num_items),
cub::Max(),
// replace with
// std::numeric_limits<T>::lowest()
// when C++11 support is more
// prevalent
Traits<InitT>::Lowest(),
stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t
Max(void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return Max<InputIteratorT, OutputIteratorT>(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_RUNTIME_FUNCTION static cudaError_t ArgMax(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::ArgMax");
// Signed integer type for global offsets
using OffsetT = int;
// The input type
using InputValueT = cub::detail::value_t<InputIteratorT>;
// The output tuple type
using OutputTupleT = cub::detail::non_void_value_t<OutputIteratorT, KeyValuePair<OffsetT, InputValueT>>;
using AccumT = OutputTupleT;
// The output value type
using OutputValueT = typename OutputTupleT::Value;
using InitT = detail::reduce::empty_problem_init_t<AccumT>;
// Wrapped input iterator to produce index-value <OffsetT, InputT> tuples
using ArgIndexInputIteratorT = ArgIndexInputIterator<InputIteratorT, OffsetT, OutputValueT>;
ArgIndexInputIteratorT d_indexed_in(d_in);
// Initial value
// TODO Address https://github.com/NVIDIA/cub/issues/651
InitT initial_value{AccumT(1, Traits<InputValueT>::Lowest())};
return DispatchReduce<ArgIndexInputIteratorT, OutputIteratorT, OffsetT, cub::ArgMax, InitT, AccumT>::Dispatch(
d_temp_storage, temp_storage_bytes, d_indexed_in, d_out, num_items, cub::ArgMax(), initial_value, stream);
}
template <typename InputIteratorT, typename OutputIteratorT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t ArgMax(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return ArgMax<InputIteratorT, OutputIteratorT>(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream);
}
template <typename InputIteratorT,
typename OutputIteratorT,
typename ReductionOpT,
typename TransformOpT,
typename T,
typename NumItemsT>
CUB_RUNTIME_FUNCTION static cudaError_t TransformReduce(
void* d_temp_storage,
size_t& temp_storage_bytes,
InputIteratorT d_in,
OutputIteratorT d_out,
NumItemsT num_items,
ReductionOpT reduction_op,
TransformOpT transform_op,
T init,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::TransformReduce");
using OffsetT = detail::choose_offset_t<NumItemsT>;
return DispatchTransformReduce<InputIteratorT, OutputIteratorT, OffsetT, ReductionOpT, TransformOpT, T>::Dispatch(
d_temp_storage,
temp_storage_bytes,
d_in,
d_out,
static_cast<OffsetT>(num_items),
reduction_op,
init,
stream,
transform_op);
}
template <typename KeysInputIteratorT,
typename UniqueOutputIteratorT,
typename ValuesInputIteratorT,
typename AggregatesOutputIteratorT,
typename NumRunsOutputIteratorT,
typename ReductionOpT,
typename NumItemsT>
CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE static cudaError_t ReduceByKey(
void* d_temp_storage,
size_t& temp_storage_bytes,
KeysInputIteratorT d_keys_in,
UniqueOutputIteratorT d_unique_out,
ValuesInputIteratorT d_values_in,
AggregatesOutputIteratorT d_aggregates_out,
NumRunsOutputIteratorT d_num_runs_out,
ReductionOpT reduction_op,
NumItemsT num_items,
cudaStream_t stream = 0)
{
CUB_DETAIL_NVTX_RANGE_SCOPE_IF(d_temp_storage, "cub::DeviceReduce::ReduceByKey");
// Signed integer type for global offsets
using OffsetT = detail::choose_offset_t<NumItemsT>;
// FlagT iterator type (not used)
// Selection op (not used)
// Default == operator
using EqualityOp = Equality;
return DispatchReduceByKey<
KeysInputIteratorT,
UniqueOutputIteratorT,
ValuesInputIteratorT,
AggregatesOutputIteratorT,
NumRunsOutputIteratorT,
EqualityOp,
ReductionOpT,
OffsetT>::Dispatch(d_temp_storage,
temp_storage_bytes,
d_keys_in,
d_unique_out,
d_values_in,
d_aggregates_out,
d_num_runs_out,
EqualityOp(),
reduction_op,
static_cast<OffsetT>(num_items),
stream);
}
template <typename KeysInputIteratorT,
typename UniqueOutputIteratorT,
typename ValuesInputIteratorT,
typename AggregatesOutputIteratorT,
typename NumRunsOutputIteratorT,
typename ReductionOpT>
CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE static cudaError_t ReduceByKey(
void* d_temp_storage,
size_t& temp_storage_bytes,
KeysInputIteratorT d_keys_in,
UniqueOutputIteratorT d_unique_out,
ValuesInputIteratorT d_values_in,
AggregatesOutputIteratorT d_aggregates_out,
NumRunsOutputIteratorT d_num_runs_out,
ReductionOpT reduction_op,
int num_items,
cudaStream_t stream,
bool debug_synchronous)
{
CUB_DETAIL_RUNTIME_DEBUG_SYNC_USAGE_LOG
return ReduceByKey<KeysInputIteratorT,
UniqueOutputIteratorT,
ValuesInputIteratorT,
AggregatesOutputIteratorT,
NumRunsOutputIteratorT,
ReductionOpT>(
d_temp_storage,
temp_storage_bytes,
d_keys_in,
d_unique_out,
d_values_in,
d_aggregates_out,
d_num_runs_out,
reduction_op,
num_items,
stream);
}
};
CUB_NAMESPACE_END