CUDASTDExecutionPolicy Include the Header CUDASTDExecutionPolicy_1CUDASTDExecutionPolicyIncludeTheHeader Parameterize Performance CUDASTDExecutionPolicy_1CUDASTDParameterizePerformance Define an Execution Policy CUDASTDExecutionPolicy_1CUDASTDDefineAnExecutionPolicy Allocate Memory Buffer for Algorithms CUDASTDExecutionPolicy_1CUDASTDAllocateMemoryBufferForAlgorithms Taskflow provides standalone template methods for expressing common parallel algorithms on a GPU. Each of these methods is governed by an execution policy object to configure the kernel execution parameters. You need to include the header file, taskflow/cuda/cudaflow.hpp, for creating a CUDA execution policy object. #include<taskflow/cuda/cudaflow.hpp> Taskflow parameterizes most CUDA algorithms in terms of the number of threads per block and units of work per thread, which can be specified in the execution policy template type, tf::cudaExecutionPolicy. The design is inspired by Modern GPU Programming authored by Sean Baxter to achieve high-performance GPU computing. The following example defines an execution policy object, policy, which configures (1) each block to invoke 512 threads and (2) each of these 512 threads to perform 11 units of work. Block size must be a power of two. It is always a good idea to specify an odd number in the second parameter to avoid bank conflicts. tf::cudaExecutionPolicy<512, 11>policy; By default, the execution policy object is associated with the CUDA default stream (i.e., 0). Default stream can incur significant overhead due to the global synchronization. You can associate an execution policy with another stream as shown below: //createaRAII-styledstreamobject tf::cudaStreamstream1,stream2; //assignastreamtoapolicyatconstructiontime tf::cudaExecutionPolicy<512, 11>policy(stream1); //assignanotherstreamtothepolicy policy.stream(stream2); All the CUDA standard algorithms in Taskflow are asynchronous with respect to the stream assigned to the execution policy. This enables high execution efficiency for large GPU workloads that call for many different algorithms. You can synchronize the stream the block until all tasks in the stream finish: cudaStreamSynchronize(policy.stream()); The best-performing configurations for each algorithm, each GPU architecture, and each data type can vary significantly. You should experiment different configurations and find the optimal tuning parameters for your applications. A default policy is given in tf::cudaDefaultExecutionPolicy. tf::cudaDefaultExecutionPolicydefault_policy; A key difference between our CUDA standard algorithms and others (e.g., Thrust) is the memory management. Unlike CPU-parallel algorithms, many GPU-parallel algorithms require extra buffer to store the temporary results during the multi-phase computation, for instance, tf::cuda_reduce and tf::cuda_sort. We DO NOT allocate any memory during these algorithms call but ask you to provide the memory buffer required for each of such algorithms. This decision seems to complicate the code a little bit, but it gives applications freedom to optimize the memory; also, it makes all algorithm calls capturable to a CUDA graph to improve the execution efficiency.