Store#

Class#

template<class T, unsigned int ItemsPerThread, unsigned int WarpSize = ::rocprim::device_warp_size(), warp_store_method Method = warp_store_method::warp_store_direct>
class warp_store#

The warp_store class is a warp level parallel primitive which provides methods for storing an arrangement of items into a blocked/striped arrangement on continous memory.

Overview

Example:

In the example a store operation is performed on a warp of 8 threads, using type int and 4 items per thread.

__global__ void example_kernel(int * output, ...)
{
    constexpr unsigned int threads_per_block = 128;
    constexpr unsigned int threads_per_warp  =   8;
    constexpr unsigned int items_per_thread  =   4;
    constexpr unsigned int warps_per_block   = threads_per_block / threads_per_warp;
    const unsigned int warp_id = hipThreadIdx_x / threads_per_warp;
    const int offset = blockIdx.x * threads_per_block * items_per_thread
        + warp_id * threads_per_warp * items_per_thread;
    int items[items_per_thread];
    rocprim::warp_store<int, items_per_thread, threads_per_warp, load_method> warp_store;
    warp_store.store(output + offset, items);
    ...
}

Template Parameters:
  • T – - the output/output type.

  • ItemsPerThread – - the number of items to be processed by each thread.

  • WarpSize – - the number of threads in a warp. It must be a divisor of the kernel block size.

  • Method – - the method to store data.

Public Types

using storage_type = storage_type_#

Struct used to allocate a temporary memory that is required for thread communication during operations provided by related parallel primitive.

Depending on the implemention the operations exposed by parallel primitive may require a temporary storage for thread communication. The storage should be allocated using keywords . It can be aliased to an externally allocated memory, or be a part of a union with other storage types to increase shared memory reusability.

Public Functions

template<class OutputIterator>
__device__ inline void store(OutputIterator output, T (&items)[ItemsPerThread], storage_type&)#

Stores an arrangement of items from across the warp into an arrangement on continuous memory.

Overview

  • The type T must be such that an object of type OutputIterator can be dereferenced and then implicitly assigned from T.

Storage reusage

Synchronization barrier should be placed before storage is reused or repurposed: __syncthreads() or rocprim::syncthreads().

Template Parameters:

OutputIterator – - [inferred] an iterator type for output (can be a simple pointer.

Parameters:
  • output[out] - the output iterator to store to.

  • items[in] - array that data is read from.

template<class OutputIterator>
__device__ inline void store(OutputIterator output, T (&items)[ItemsPerThread], unsigned int valid, storage_type&)#

Stores an arrangement of items from across the warp into an arrangement on continuous memory, which is guarded by range valid, using temporary storage.

Overview

  • The type T must be such that an object of type OutputIterator can be dereferenced and then implicitly assigned from T.

Storage reusage

Synchronization barrier should be placed before storage is reused or repurposed: __syncthreads() or rocprim::syncthreads().

Template Parameters:

OutputIterator – - [inferred] an iterator type for output (can be a simple pointer.

Parameters:
  • output[out] - the output iterator to store to.

  • items[in] - array that data is read from.

  • valid[in] - maximum range of valid numbers to read.

Algorithms#

enum class rocprim::warp_store_method#

warp_store_method enumerates the methods available to store a blocked/striped arrangement of items into a blocked/striped arrangement in continuous memory

Values:

enumerator warp_store_direct#

A blocked arrangement of items is stored into a blocked arrangement on continuous memory.

Performance Notes:

  • Performance decreases with increasing number of items per thread (stride between reads), because of reduced memory coalescing.

enumerator warp_store_striped#

A striped arrangement of items is stored into a blocked arrangement on continuous memory.

enumerator warp_store_vectorize#

A blocked arrangement of items is stored into a blocked arrangement on continuous memory using vectorization as an optimization.

Performance Notes:

  • Performance remains high due to increased memory coalescing, provided that vectorization requirements are fulfilled. Otherwise, performance will default to warp_store_direct.

Requirements:

  • The output offset (block_output) must be quad-item aligned.

  • The following conditions will prevent vectorization and switch to default warp_store_direct:

    • ItemsPerThread is odd.

    • The datatype T is not a primitive or a HIP vector type (e.g. int2, int4, etc.

enumerator warp_store_transpose#

A blocked arrangement of items is locally transposed and stored as a striped arrangement of data on continuous memory.

Performance Notes:

  • Performance remains high due to increased memory coalescing, regardless of the number of items per thread.

  • Performance may be better compared to warp_store_direct and warp_store_vectorize due to reordering on local memory.

enumerator default_method#

Defaults to warp_store_direct.