CUB  
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
List of all members
cub::DeviceScan Struct Reference
Single-problem

Detailed description

DeviceScan provides device-wide, parallel operations for computing a prefix scan across a sequence of data items residing within device-accessible memory.

device_scan.png
.
Overview
Given a sequence of input elements and a binary reduction operator, a prefix scan produces an output sequence where each element is computed to be the reduction of the elements occurring earlier in the input sequence. Prefix sum connotes a prefix scan with the addition operator. The term inclusive indicates that the ith output reduction incorporates the ith input. The term exclusive indicates the ith input is not incorporated into the ith output reduction.
As of CUB 1.0.1 (2013), CUB's device-wide scan APIs have implemented our "decoupled look-back" algorithm for performing global prefix scan with only a single pass through the input data, as described in our 2016 technical report [1]. The central idea is to leverage a small, constant factor of redundant work in order to overlap the latencies of global prefix propagation with local computation. As such, our algorithm requires only ~2n data movement (n inputs are read, n outputs are written), and typically proceeds at "memcpy" speeds. Our algorithm supports inplace operations.
[1] Duane Merrill and Michael Garland. "Single-pass Parallel Prefix Scan with Decoupled Look-back", NVIDIA Technical Report NVR-2016-002, 2016.
Usage Considerations
  • Dynamic parallelism. DeviceScan methods can be called within kernel code on devices in which CUDA dynamic parallelism is supported.
Performance
The work-complexity of prefix scan 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 DeviceScan::ExclusiveSum performance across different CUDA architectures for int32 keys. Performance plots for other scenarios can be found in the detailed method descriptions below.
scan_int32.png

Static Public Methods

Exclusive scans
template<typename InputIteratorT , typename OutputIteratorT >
static CUB_RUNTIME_FUNCTION
cudaError_t 		ExclusiveSum (void *d_temp_storage, size_t &temp_storage_bytes, InputIteratorT d_in, OutputIteratorT d_out, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide exclusive prefix sum. The value of 0 is applied as the initial value, and is assigned to *d_out. More...
 
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT , typename InitValueT >
static CUB_RUNTIME_FUNCTION
cudaError_t 		ExclusiveScan (void *d_temp_storage, size_t &temp_storage_bytes, InputIteratorT d_in, OutputIteratorT d_out, ScanOpT scan_op, InitValueT init_value, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide exclusive prefix scan using the specified binary scan_op functor. The init_value value is applied as the initial value, and is assigned to *d_out. More...
 
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT , typename InitValueT , typename InitValueIterT = InitValueT*>
static CUB_RUNTIME_FUNCTION
cudaError_t 		ExclusiveScan (void *d_temp_storage, size_t &temp_storage_bytes, InputIteratorT d_in, OutputIteratorT d_out, ScanOpT scan_op, FutureValue< InitValueT, InitValueIterT > init_value, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 
Inclusive scans
template<typename InputIteratorT , typename OutputIteratorT >
static CUB_RUNTIME_FUNCTION
cudaError_t 		InclusiveSum (void *d_temp_storage, size_t &temp_storage_bytes, InputIteratorT d_in, OutputIteratorT d_out, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide inclusive prefix sum. More...
 
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT >
static CUB_RUNTIME_FUNCTION
cudaError_t 		InclusiveScan (void *d_temp_storage, size_t &temp_storage_bytes, InputIteratorT d_in, OutputIteratorT d_out, ScanOpT scan_op, int num_items, cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide inclusive prefix scan using the specified binary scan_op functor. More...
 
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION
cudaError_t 		ExclusiveSumByKey (void *d_temp_storage, size_t &temp_storage_bytes, KeysInputIteratorT d_keys_in, ValuesInputIteratorT d_values_in, ValuesOutputIteratorT d_values_out, int num_items, EqualityOpT equality_op=EqualityOpT(), cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide exclusive prefix sum-by-key with key equality defined by equality_op . The value of 0 is applied as the initial value, and is assigned to the beginning of each segment in d_values_out . More...
 
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename ScanOpT , typename InitValueT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION
cudaError_t 		ExclusiveScanByKey (void *d_temp_storage, size_t &temp_storage_bytes, KeysInputIteratorT d_keys_in, ValuesInputIteratorT d_values_in, ValuesOutputIteratorT d_values_out, ScanOpT scan_op, InitValueT init_value, int num_items, EqualityOpT equality_op=EqualityOpT(), cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide exclusive prefix scan-by-key using the specified binary scan_op functor. The key equality is defined by equality_op . The init_value value is applied as the initial value, and is assigned to the beginning of each segment in d_values_out . More...
 
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION
cudaError_t 		InclusiveSumByKey (void *d_temp_storage, size_t &temp_storage_bytes, KeysInputIteratorT d_keys_in, ValuesInputIteratorT d_values_in, ValuesOutputIteratorT d_values_out, int num_items, EqualityOpT equality_op=EqualityOpT(), cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide inclusive prefix sum-by-key with key equality defined by equality_op . More...
 
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename ScanOpT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION
cudaError_t 		InclusiveScanByKey (void *d_temp_storage, size_t &temp_storage_bytes, KeysInputIteratorT d_keys_in, ValuesInputIteratorT d_values_in, ValuesOutputIteratorT d_values_out, ScanOpT scan_op, int num_items, EqualityOpT equality_op=EqualityOpT(), cudaStream_t stream=0, bool debug_synchronous=false)
 Computes a device-wide inclusive prefix scan-by-key using the specified binary scan_op functor. The key equality is defined by equality_op . More...
 

Member Function Documentation

template<typename InputIteratorT , typename OutputIteratorT >
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::ExclusiveSum ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIteratorT  d_in,
OutputIteratorT  d_out,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide exclusive prefix sum. The value of 0 is applied as the initial value, and is assigned to *d_out.

  • Supports non-commutative sum operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Performance
The following charts illustrate saturated exclusive sum performance across different CUDA architectures for int32 and int64 items, respectively.
scan_int32.png
scan_int64.png
Snippet
The code snippet below illustrates the exclusive prefix sum of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_out; // e.g., [ , , , , , , ]
...
// Determine temporary device storage requirements
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
// Allocate temporary storage
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run exclusive prefix sum
cub::DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
// d_out s<-- [0, 8, 14, 21, 26, 29, 29]
Template Parameters
InputIteratorT[inferred] Random-access input iterator type for reading scan inputs (may be a simple pointer type)
OutputIteratorT[inferred] Random-access output iterator type for writing scan outputs (may be a simple pointer type)
Parameters
[in]d_temp_storageDevice-accessible 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_inRandom-access iterator to the input sequence of data items
[out]d_outRandom-access iterator to the output sequence of data items
[in]num_itemsTotal number of input items (i.e., the length of d_in)
[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_scan.cu.
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT , typename InitValueT >
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::ExclusiveScan ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIteratorT  d_in,
OutputIteratorT  d_out,
ScanOpT  scan_op,
InitValueT  init_value,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide exclusive prefix scan using the specified binary scan_op functor. The init_value value is applied as the initial value, and is assigned to *d_out.

  • Supports non-commutative scan operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the exclusive prefix min-scan of an int device vector
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
#include <climits> // for INT_MAX
// CustomMin functor
struct CustomMin
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return (b < a) ? b : a;
}
};
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_out; // e.g., [ , , , , , , ]
CustomMin min_op;
...
// Determine temporary device storage requirements for exclusive prefix scan
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::ExclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, (int) INT_MAX, num_items);
// Allocate temporary storage for exclusive prefix scan
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run exclusive prefix min-scan
cub::DeviceScan::ExclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, (int) INT_MAX, num_items);
// d_out <-- [2147483647, 8, 6, 6, 5, 3, 0]
Template Parameters
InputIteratorT[inferred] Random-access input iterator type for reading scan inputs (may be a simple pointer type)
OutputIteratorT[inferred] Random-access output iterator type for writing scan outputs (may be a simple pointer type)
ScanOp[inferred] Binary scan functor type having member T operator()(const T &a, const T &b)
InitValueT[inferred] Type of the init_value used Binary scan functor type having member T operator()(const T &a, const T &b)
Parameters
[in]d_temp_storageDevice-accessible 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_inRandom-access iterator to the input sequence of data items
[out]d_outRandom-access iterator to the output sequence of data items
[in]scan_opBinary scan functor
[in]init_valueInitial value to seed the exclusive scan (and is assigned to *d_out)
[in]num_itemsTotal number of input items (i.e., the length of d_in)
[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.
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT , typename InitValueT , typename InitValueIterT = InitValueT*>
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::ExclusiveScan ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIteratorT  d_in,
OutputIteratorT  d_out,
ScanOpT  scan_op,
FutureValue< InitValueT, InitValueIterT >  init_value,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic
Parameters
[in]d_temp_storageDevice-accessible 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 data items
[in]scan_opBinary scan functor
[in]init_valueInitial value to seed the exclusive scan (and is assigned to *d_out)
[in]num_itemsTotal number of input items (i.e., the length of d_in)
[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.
template<typename InputIteratorT , typename OutputIteratorT >
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::InclusiveSum ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIteratorT  d_in,
OutputIteratorT  d_out,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide inclusive prefix sum.

  • Supports non-commutative sum operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the inclusive prefix sum of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_out; // e.g., [ , , , , , , ]
...
// Determine temporary device storage requirements for inclusive prefix sum
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
// Allocate temporary storage for inclusive prefix sum
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run inclusive prefix sum
cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
// d_out <-- [8, 14, 21, 26, 29, 29, 38]
Template Parameters
InputIteratorT[inferred] Random-access input iterator type for reading scan inputs (may be a simple pointer type)
OutputIteratorT[inferred] Random-access output iterator type for writing scan outputs (may be a simple pointer type)
Parameters
[in]d_temp_storageDevice-accessible 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_inRandom-access iterator to the input sequence of data items
[out]d_outRandom-access iterator to the output sequence of data items
[in]num_itemsTotal number of input items (i.e., the length of d_in)
[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.
template<typename InputIteratorT , typename OutputIteratorT , typename ScanOpT >
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::InclusiveScan ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
InputIteratorT  d_in,
OutputIteratorT  d_out,
ScanOpT  scan_op,
int  num_items,
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide inclusive prefix scan using the specified binary scan_op functor.

  • Supports non-commutative scan operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the inclusive prefix min-scan of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
#include <climits> // for INT_MAX
// CustomMin functor
struct CustomMin
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return (b < a) ? b : a;
}
};
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_out; // e.g., [ , , , , , , ]
CustomMin min_op;
...
// Determine temporary device storage requirements for inclusive prefix scan
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::InclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, num_items);
// Allocate temporary storage for inclusive prefix scan
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run inclusive prefix min-scan
cub::DeviceScan::InclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, num_items);
// d_out <-- [8, 6, 6, 5, 3, 0, 0]
Template Parameters
InputIteratorT[inferred] Random-access input iterator type for reading scan inputs (may be a simple pointer type)
OutputIteratorT[inferred] Random-access output iterator type for writing scan outputs (may be a simple pointer type)
ScanOp[inferred] Binary scan functor type having member T operator()(const T &a, const T &b)
Parameters
[in]d_temp_storageDevice-accessible 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_inRandom-access iterator to the input sequence of data items
[out]d_outRandom-access iterator to the output sequence of data items
[in]scan_opBinary scan functor
[in]num_itemsTotal number of input items (i.e., the length of d_in)
[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.
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::ExclusiveSumByKey ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
KeysInputIteratorT  d_keys_in,
ValuesInputIteratorT  d_values_in,
ValuesOutputIteratorT  d_values_out,
int  num_items,
EqualityOpT  equality_op = EqualityOpT(),
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide exclusive prefix sum-by-key with key equality defined by equality_op . The value of 0 is applied as the initial value, and is assigned to the beginning of each segment in d_values_out .

  • Supports non-commutative sum operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the exclusive prefix sum-by-key of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_keys_in; // e.g., [0, 0, 1, 1, 1, 2, 2]
int *d_values_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_values_out; // e.g., [ , , , , , , ]
...
// Determine temporary device storage requirements
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::ExclusiveSumByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, num_items);
// Allocate temporary storage
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run exclusive prefix sum
cub::DeviceScan::ExclusiveSumByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, num_items);
// d_values_out <-- [0, 8, 0, 7, 12, 0, 0]
Template Parameters
KeysInputIteratorT[inferred] Random-access input iterator type for reading scan keys inputs (may be a simple pointer type)
ValuesInputIteratorT[inferred] Random-access input iterator type for reading scan values inputs (may be a simple pointer type)
ValuesOutputIteratorT[inferred] Random-access output iterator type for writing scan values outputs (may be a simple pointer type)
EqualityOpT[inferred] Functor type having member T operator()(const T &a, const T &b) for binary operations that defines the equality of keys
Parameters
[in]d_temp_storageDevice-accessible 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_keys_inRandom-access input iterator to the input sequence of key items
[in]d_values_inRandom-access input iterator to the input sequence of value items
[out]d_values_outRandom-access output iterator to the output sequence of value items
[in]num_itemsTotal number of input items (i.e., the length of d_keys_in and d_values_in)
[in]equality_opBinary functor that defines the equality of keys. Default is cub::Equality().
[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.
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename ScanOpT , typename InitValueT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::ExclusiveScanByKey ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
KeysInputIteratorT  d_keys_in,
ValuesInputIteratorT  d_values_in,
ValuesOutputIteratorT  d_values_out,
ScanOpT  scan_op,
InitValueT  init_value,
int  num_items,
EqualityOpT  equality_op = EqualityOpT(),
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide exclusive prefix scan-by-key using the specified binary scan_op functor. The key equality is defined by equality_op . The init_value value is applied as the initial value, and is assigned to the beginning of each segment in d_values_out .

  • Supports non-commutative scan operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the exclusive prefix min-scan-by-key of an int device vector
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
#include <climits> // for INT_MAX
// CustomMin functor
struct CustomMin
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return (b < a) ? b : a;
}
};
// CustomEqual functor
struct CustomEqual
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return a == b;
}
};
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_keys_in; // e.g., [0, 0, 1, 1, 1, 2, 2]
int *d_values_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_values_out; // e.g., [ , , , , , , ]
CustomMin min_op;
CustomEqual equality_op;
...
// Determine temporary device storage requirements for exclusive prefix scan
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::ExclusiveScanByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, min_op, (int) INT_MAX, num_items, equality_op);
// Allocate temporary storage for exclusive prefix scan
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run exclusive prefix min-scan
cub::DeviceScan::ExclusiveScanByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, min_op, (int) INT_MAX, num_items, equality_op);
// d_values_out <-- [2147483647, 8, 2147483647, 7, 5, 2147483647, 0]
Template Parameters
KeysInputIteratorT[inferred] Random-access input iterator type for reading scan keys inputs (may be a simple pointer type)
ValuesInputIteratorT[inferred] Random-access input iterator type for reading scan values inputs (may be a simple pointer type)
ValuesOutputIteratorT[inferred] Random-access output iterator type for writing scan values outputs (may be a simple pointer type)
ScanOp[inferred] Binary scan functor type having member T operator()(const T &a, const T &b)
InitValueT[inferred] Type of the init_value value used in Binary scan functor type having member T operator()(const T &a, const T &b)
EqualityOpT[inferred] Functor type having member T operator()(const T &a, const T &b) for binary operations that defines the equality of keys
Parameters
[in]d_temp_storageDevice-accessible 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_keys_inRandom-access input iterator to the input sequence of key items
[in]d_values_inRandom-access input iterator to the input sequence of value items
[out]d_values_outRandom-access output iterator to the output sequence of value items
[in]scan_opBinary scan functor
[in]init_valueInitial value to seed the exclusive scan (and is assigned to the beginning of each segment in d_values_out)
[in]num_itemsTotal number of input items (i.e., the length of d_keys_in and d_values_in)
[in]equality_opBinary functor that defines the equality of keys. Default is cub::Equality().
[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.
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::InclusiveSumByKey ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
KeysInputIteratorT  d_keys_in,
ValuesInputIteratorT  d_values_in,
ValuesOutputIteratorT  d_values_out,
int  num_items,
EqualityOpT  equality_op = EqualityOpT(),
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide inclusive prefix sum-by-key with key equality defined by equality_op .

  • Supports non-commutative sum operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the inclusive prefix sum-by-key of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_keys_in; // e.g., [0, 0, 1, 1, 1, 2, 2]
int *d_values_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_values_out; // e.g., [ , , , , , , ]
...
// Determine temporary device storage requirements for inclusive prefix sum
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::InclusiveSumByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, num_items);
// Allocate temporary storage for inclusive prefix sum
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run inclusive prefix sum
cub::DeviceScan::InclusiveSumByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, num_items);
// d_out <-- [8, 14, 7, 12, 15, 0, 9]
Template Parameters
KeysInputIteratorT[inferred] Random-access input iterator type for reading scan keys inputs (may be a simple pointer type)
ValuesInputIteratorT[inferred] Random-access input iterator type for reading scan values inputs (may be a simple pointer type)
ValuesOutputIteratorT[inferred] Random-access output iterator type for writing scan values outputs (may be a simple pointer type)
EqualityOpT[inferred] Functor type having member T operator()(const T &a, const T &b) for binary operations that defines the equality of keys
Parameters
[in]d_temp_storageDevice-accessible 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_keys_inRandom-access input iterator to the input sequence of key items
[in]d_values_inRandom-access input iterator to the input sequence of value items
[out]d_values_outRandom-access output iterator to the output sequence of value items
[in]num_itemsTotal number of input items (i.e., the length of d_keys_in and d_values_in)
[in]equality_opBinary functor that defines the equality of keys. Default is cub::Equality().
[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.
template<typename KeysInputIteratorT , typename ValuesInputIteratorT , typename ValuesOutputIteratorT , typename ScanOpT , typename EqualityOpT = Equality>
static CUB_RUNTIME_FUNCTION cudaError_t cub::DeviceScan::InclusiveScanByKey ( void *  d_temp_storage,
size_t &  temp_storage_bytes,
KeysInputIteratorT  d_keys_in,
ValuesInputIteratorT  d_values_in,
ValuesOutputIteratorT  d_values_out,
ScanOpT  scan_op,
int  num_items,
EqualityOpT  equality_op = EqualityOpT(),
cudaStream_t  stream = 0,
bool  debug_synchronous = false 
)
inlinestatic

Computes a device-wide inclusive prefix scan-by-key using the specified binary scan_op functor. The key equality is defined by equality_op .

  • Supports non-commutative scan operators.
  • Results are not deterministic for pseudo-associative operators (e.g., addition of floating-point types). Results for pseudo-associative operators may vary from run to run. Additional details can be found in the decoupled look-back description.
  • When d_temp_storage is NULL, no work is done and the required allocation size is returned in temp_storage_bytes.
Snippet
The code snippet below illustrates the inclusive prefix min-scan-by-key of an int device vector.
#include <cub/cub.cuh> // or equivalently <cub/device/device_scan.cuh>
#include <climits> // for INT_MAX
// CustomMin functor
struct CustomMin
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return (b < a) ? b : a;
}
};
// CustomEqual functor
struct CustomEqual
{
template <typename T>
CUB_RUNTIME_FUNCTION __forceinline__
T operator()(const T &a, const T &b) const {
return a == b;
}
};
// Declare, allocate, and initialize device-accessible pointers for input and output
int num_items; // e.g., 7
int *d_keys_in; // e.g., [0, 0, 1, 1, 1, 2, 2]
int *d_values_in; // e.g., [8, 6, 7, 5, 3, 0, 9]
int *d_values_out; // e.g., [ , , , , , , ]
CustomMin min_op;
CustomEqual equality_op;
...
// Determine temporary device storage requirements for inclusive prefix scan
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceScan::InclusiveScanByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, min_op, num_items, equality_op);
// Allocate temporary storage for inclusive prefix scan
cudaMalloc(&d_temp_storage, temp_storage_bytes);
// Run inclusive prefix min-scan
cub::DeviceScan::InclusiveScanByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_values_in, d_values_out, min_op, num_items, equality_op);
// d_out <-- [8, 6, 7, 5, 3, 0, 0]
Template Parameters
KeysInputIteratorT[inferred] Random-access input iterator type for reading scan keys inputs (may be a simple pointer type)
ValuesInputIteratorT[inferred] Random-access input iterator type for reading scan values inputs (may be a simple pointer type)
ValuesOutputIteratorT[inferred] Random-access output iterator type for writing scan values outputs (may be a simple pointer type)
ScanOp[inferred] Binary scan functor type having member T operator()(const T &a, const T &b)
EqualityOpT[inferred] Functor type having member T operator()(const T &a, const T &b) for binary operations that defines the equality of keys
Parameters
[in]d_temp_storageDevice-accessible 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_keys_inRandom-access input iterator to the input sequence of key items
[in]d_values_inRandom-access input iterator to the input sequence of value items
[out]d_values_outRandom-access output iterator to the output sequence of value items
[in]scan_opBinary scan functor
[in]num_itemsTotal number of input items (i.e., the length of d_keys_in and d_values_in)
[in]equality_opBinary functor that defines the equality of keys. Default is cub::Equality().
[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.
The documentation for this struct was generated from the following file:
Generated by   doxygen 1.8.4
© 2022 NVIDIA Corporation