CUB
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Groups
Static Public Methods | List of all members
cub::DevicePartition Struct Reference

Detailed description

DevicePartition provides device-wide, parallel operations for partitioning sequences of data items residing within global memory.

partition_logo.png
.
Overview
These operations apply a selection criterion to construct a partitioned output sequence from items selected/unselected from a specified input sequence.
Usage Considerations
  • Dynamic parallelism. DevicePartition methods can be called within kernel code on devices in which CUDA dynamic parallelism is supported. When calling these methods from kernel code, be sure to define the CUB_CDP macro in your compiler's macro definitions.
Performance
The work-complexity of partition as a function of input size is linear, resulting in performance throughput that plateaus with problem sizes large enough to saturate the GPU.
The following chart illustrates DevicePartition::If performance across different CUDA architectures for int32 items, where 50% of the items are randomly selected for the first partition. Performance plots for other scenarios can be found in the detailed method descriptions below.
partition_if_int32_50_percent.png

Definition at line 73 of file device_partition.cuh.

Static Public Methods

template<typename InputIterator , typename FlagIterator , typename OutputIterator , typename NumSelectedIterator >
CUB_RUNTIME_FUNCTION static
__forceinline__ cudaError_t 
Flagged (void *d_temp_storage, size_t &temp_storage_bytes, InputIterator d_in, FlagIterator d_flags, OutputIterator d_out, NumSelectedIterator d_num_selected, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 Uses the d_flags sequence to split the corresponding items from d_in into a partitioned sequence d_out. The total number of items copied into the first partition is written to d_num_selected.

partition_flags_logo.png
.
More...
 
template<typename InputIterator , typename OutputIterator , typename NumSelectedIterator , typename SelectOp >
CUB_RUNTIME_FUNCTION static
__forceinline__ cudaError_t 
If (void *d_temp_storage, size_t &temp_storage_bytes, InputIterator d_in, OutputIterator d_out, NumSelectedIterator d_num_selected, int num_items, SelectOp select_op, cudaStream_t stream=0, bool debug_synchronous=false)
 Uses the select_op functor to split the corresponding items from d_in into a partitioned sequence d_out. The total number of items copied into the first partition is written to d_num_selected.

partition_logo.png
.
More...
 

Member Function Documentation

template<typename InputIterator , typename FlagIterator , typename OutputIterator , typename NumSelectedIterator >
CUB_RUNTIME_FUNCTION static __forceinline__ cudaError_t cub::DevicePartition::Flagged ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIterator  d_in,
FlagIterator  d_flags,
OutputIterator  d_out,
NumSelectedIterator  d_num_selected,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Uses the d_flags sequence to split the corresponding items from d_in into a partitioned sequence d_out. The total number of items copied into the first partition is written to d_num_selected.

partition_flags_logo.png
.

  • The value type of d_flags must be castable to bool (e.g., bool, char, int, etc.).
  • Copies of the selected items are compacted into d_out and maintain their original relative ordering, however copies of the unselected items are compacted into the rear of d_out in reverse order.
  • This operation requires an allocation of temporary device storage. When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
  • When calling this method from kernel code, be sure to define the CUB_CDP macro in your compiler's macro definitions.
Snippet
The code snippet below illustrates the compaction of items selected from an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_partition.cuh>
// Declare, allocate, and initialize device pointers for input, flags, and output
int num_items; // e.g., 8
int *d_in; // e.g., [1, 2, 3, 4, 5, 6, 7, 8]
char *d_flags; // e.g., [1, 0, 0, 1, 0, 1, 1, 0]
int *d_out; // e.g., [ , , , , , , , ]
int *d_num_selected; // e.g., [ ]
...
// Determine temporary device storage requirements
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected, num_items);
// Allocate temporary storage
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run selection
cub::DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected, num_items);
// d_out <-- [1, 4, 6, 7, 8, 5, 3, 2]
// d_num_selected <-- [4]
Template Parameters
InputIterator[inferred] Random-access input iterator type for reading input items (may be a simple pointer type)
FlagIterator[inferred] Random-access input iterator type for reading selection flags (may be a simple pointer type)
OutputIterator[inferred] Random-access output iterator type for writing output items (may be a simple pointer type)
NumSelectedIterator[inferred] Output iterator type for recording the number of items selected (may be a simple pointer type)
Parameters
[in]d_temp_storageDevice allocation of temporary storage. When NULL, the required allocation size is written to temp_storage_bytes and no work is done.
[in,out]temp_storage_bytesReference to size in bytes of d_temp_storage allocation
[in]d_inPointer to the input sequence of data items
[in]d_flagsPointer to the input sequence of selection flags
[out]d_outPointer to the output sequence of partitioned data items
[out]d_num_selectedPointer to the output total number of items selected (i.e., the offset of the unselected partition)
[in]num_itemsTotal number of items to select from
[in]stream[optional] CUDA stream to launch kernels within. Default is stream0.
[in]debug_synchronous[optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is false.
Examples:
example_device_partition_flagged.cu.

Definition at line 127 of file device_partition.cuh.

template<typename InputIterator , typename OutputIterator , typename NumSelectedIterator , typename SelectOp >
CUB_RUNTIME_FUNCTION static __forceinline__ cudaError_t cub::DevicePartition::If ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIterator  d_in,
OutputIterator  d_out,
NumSelectedIterator  d_num_selected,
int  num_items,
SelectOp  select_op,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Uses the select_op functor to split the corresponding items from d_in into a partitioned sequence d_out. The total number of items copied into the first partition is written to d_num_selected.

partition_logo.png
.

  • Copies of the selected items are compacted into d_out and maintain their original relative ordering, however copies of the unselected items are compacted into the rear of d_out in reverse order.
  • This operation requires an allocation of temporary device storage. When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
  • When calling this method from kernel code, be sure to define the CUB_CDP macro in your compiler's macro definitions.
Performance
The following charts illustrate saturated partition-if performance across different CUDA architectures for int32 and int64 items, respectively. Items are selected for the first partition with 50% probability.
partition_if_int32_50_percent.png
partition_if_int64_50_percent.png
The following charts are similar, but 5% selection probability for the first partition:
partition_if_int32_5_percent.png
partition_if_int64_5_percent.png
Snippet
The code snippet below illustrates the compaction of items selected from an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_partition.cuh>
// Functor type for selecting values less than some criteria
struct LessThan
{
int compare;
CUB_RUNTIME_FUNCTION __forceinline__
LessThan(int compare) : compare(compare) {}
CUB_RUNTIME_FUNCTION __forceinline__
bool operator()(const int &a) const {
return (a < compare);
}
};
// Declare, allocate, and initialize device pointers for input and output
int num_items; // e.g., 8
int *d_in; // e.g., [0, 2, 3, 9, 5, 2, 81, 8]
int *d_out; // e.g., [ , , , , , , , ]
int *d_num_selected; // e.g., [ ]
LessThan select_op(7);
...
// Determine temporary device storage requirements
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected, num_items, select_op);
// Allocate temporary storage
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run selection
cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected, num_items, select_op);
// d_out <-- [0, 2, 3, 5, 2, 8, 81, 9]
// d_num_selected <-- [5]
Template Parameters
InputIterator[inferred] Random-access input iterator type for reading input items (may be a simple pointer type)
OutputIterator[inferred] Random-access output iterator type for writing output items (may be a simple pointer type)
NumSelectedIterator[inferred] Output iterator type for recording the number of items selected (may be a simple pointer type)
SelectOp[inferred] Selection functor type having member bool operator()(const T &a)
Parameters
[in]d_temp_storageDevice allocation of temporary storage. When NULL, the required allocation size is written to temp_storage_bytes and no work is done.
[in,out]temp_storage_bytesReference to size in bytes of d_temp_storage allocation
[in]d_inPointer to the input sequence of data items
[out]d_outPointer to the output sequence of partitioned data items
[out]d_num_selectedPointer to the output total number of items selected (i.e., the offset of the unselected partition)
[in]num_itemsTotal number of items to select from
[in]select_opUnary selection operator
[in]stream[optional] CUDA stream to launch kernels within. Default is stream0.
[in]debug_synchronous[optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is false.
Examples:
example_device_partition_if.cu.

Definition at line 236 of file device_partition.cuh.


The documentation for this struct was generated from the following file: