- Notifications
You must be signed in to change notification settings - Fork0
Cooperative primitives for CUDA C++.
License
pauleonix/cub
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
CUB provides state-of-the-art, reusable software components for every layerof the CUDA programming model:
- Device-wide primitives
- Sort, prefix scan, reduction, histogram, etc.
- Compatible with CUDA dynamic parallelism
- Block-wide "collective" primitives
- I/O, sort, prefix scan, reduction, histogram, etc.
- Compatible with arbitrary thread block sizes and types
- Warp-wide "collective" primitives
- Warp-wide prefix scan, reduction, etc.
- Safe and architecture-specific
- Thread and resource utilities
- PTX intrinsics, device reflection, texture-caching iterators, caching memory allocators, etc.
CUB is included in the NVIDIA HPC SDK and the CUDA Toolkit.
We recommend theCUB Project Website for further information and examples.
#include<cub/cub.cuh>// Block-sorting CUDA kernel__global__voidBlockSortKernel(int *d_in,int *d_out){usingnamespacecub;// Specialize BlockRadixSort, BlockLoad, and BlockStore for 128 threads// owning 16 integer items eachtypedef BlockRadixSort<int,128,16> BlockRadixSort;typedef BlockLoad<int,128,16, BLOCK_LOAD_TRANSPOSE> BlockLoad;typedef BlockStore<int,128,16, BLOCK_STORE_TRANSPOSE> BlockStore;// Allocate shared memory __shared__union {typename BlockRadixSort::TempStorage sort;typename BlockLoad::TempStorage load;typename BlockStore::TempStorage store; } temp_storage;int block_offset = blockIdx.x * (128 *16);// OffsetT for this block's ment// Obtain a segment of 2048 consecutive keys that are blocked across threadsint thread_keys[16];BlockLoad(temp_storage.load).Load(d_in + block_offset, thread_keys);__syncthreads();// Collectively sort the keysBlockRadixSort(temp_storage.sort).Sort(thread_keys);__syncthreads();// Store the sorted segmentBlockStore(temp_storage.store).Store(d_out + block_offset, thread_keys);}
Each thread block usescub::BlockRadixSort
to collectively sortits own input segment. The class is specialized by thedata type being sorted, by the number of threads per block, by the number ofkeys per thread, and implicitly by the targeted compilation architecture.
Thecub::BlockLoad
andcub::BlockStore
classes are similarly specialized.Furthermore, to provide coalesced accesses to device memory, these primitives areconfigured to access memory using a striped access pattern (where consecutive threadssimultaneously access consecutive items) and thentranspose the keys intoablocked arrangement of elements across threads.
Once specialized, these classes expose opaqueTempStorage
member types.The thread block uses these storage types to statically allocate the union ofshared memory needed by the thread block. (Alternatively these storage typescould be aliased to global memory allocations).
CUB is regularly tested using the specified versions of the followingcompilers. Unsupported versions may emit deprecation warnings, which can besilenced by defining CUB_IGNORE_DEPRECATED_COMPILER during compilation.
- NVCC 11.0+
- NVC++ 20.9+
- GCC 5+
- Clang 7+
- MSVC 2019+ (19.20/16.0/14.20)
CUB is distributed with the NVIDIA HPC SDK and the CUDA Toolkit in additionto GitHub.
See thechangelog for details about specific releases.
CUB Release | Included In |
---|---|
1.16.0 | TBD |
1.15.0 | NVIDIA HPC SDK 22.1 & CUDA Toolkit 11.6 |
1.14.0 | NVIDIA HPC SDK 21.9 |
1.13.1 | CUDA Toolkit 11.5 |
1.13.0 | NVIDIA HPC SDK 21.7 |
1.12.1 | CUDA Toolkit 11.4 |
1.12.0 | NVIDIA HPC SDK 21.3 |
1.11.0 | CUDA Toolkit 11.3 |
1.10.0 | NVIDIA HPC SDK 20.9 & CUDA Toolkit 11.2 |
1.9.10-1 | NVIDIA HPC SDK 20.7 & CUDA Toolkit 11.1 |
1.9.10 | NVIDIA HPC SDK 20.5 |
1.9.9 | CUDA Toolkit 11.0 |
1.9.8-1 | NVIDIA HPC SDK 20.3 |
1.9.8 | CUDA Toolkit 11.0 Early Access |
1.9.8 | CUDA 11.0 Early Access |
1.8.0 | |
1.7.5 | Thrust 1.9.2 |
1.7.4 | Thrust 1.9.1-2 |
1.7.3 | |
1.7.2 | |
1.7.1 | |
1.7.0 | Thrust 1.9.0-5 |
1.6.4 | |
1.6.3 | |
1.6.2 (previously 1.5.5) | |
1.6.1 (previously 1.5.4) | |
1.6.0 (previously 1.5.3) | |
1.5.2 | |
1.5.1 | |
1.5.0 | |
1.4.1 | |
1.4.0 | |
1.3.2 | |
1.3.1 | |
1.3.0 | |
1.2.3 | |
1.2.2 | |
1.2.0 | |
1.1.1 | |
1.0.2 | |
1.0.1 | |
0.9.4 | |
0.9.2 | |
0.9.1 | |
0.9.0 |
CUB and Thrust depend on each other. It is recommended to clone Thrustand build CUB as a component of Thrust.
CUB uses theCMake build system to build unit tests,examples, and header tests. To build CUB as a developer, the followingrecipe should be followed:
# Clone Thrust and CUB from Github. CUB is located in Thrust's# `dependencies/cub` submodule.git clone --recursive https://github.com/NVIDIA/thrust.gitcd thrust# Create build directory:mkdir buildcd build# Configure -- use one of the following:cmake -DTHRUST_INCLUDE_CUB_CMAKE=ON .. # Command line interface.ccmake -DTHRUST_INCLUDE_CUB_CMAKE=ON .. # ncurses GUI (Linux only)cmake-gui # Graphical UI, set source/build directories and options in the app# Build:cmake --build . -j <num jobs> # invokes make (or ninja, etc)# Run tests and examples:ctest
By default, the C++14 standard is targeted, but this can be changed in CMake.More information on configuring your CUB build and creating a pull request isfound inCONTRIBUTING.md.
CUB is available under the "New BSD" open-source license:
Copyright (c) 2010-2011, Duane Merrill. All rights reserved.Copyright (c) 2011-2018, NVIDIA CORPORATION. All rights reserved.Redistribution and use in source and binary forms, with or withoutmodification, 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" ANDANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIEDWARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE AREDISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANYDIRECT, 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 ANDON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THISSOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
About
Cooperative primitives for CUDA C++.
Resources
License
Code of conduct
Stars
Watchers
Forks
Packages0
Languages
- Cuda96.2%
- C++2.4%
- CMake1.4%