Release 3.0.0 (2021/01/01)

Table of Contents

• Download
• System Requirements
• Working Items
• New Features
  • Taskflow Core
  • cudaFlow
  • Utilities
  • Taskflow Profiler (TFProf)
• New Algorithms
  • CPU Algorithms
  • GPU Algorithms
• Bug Fixes
• Breaking Changes
• Deprecated and Removed Items
• Documentation
• Miscellaneous Items

Taskflow 3.0.0 is the 1st release in the 3.x line! This release includes several new changes such as CPU-GPU tasking, algorithm collection, enhanced web-based profiler, documentation, and unit tests.

Note

Starting from v3, we have migrated the codebase to the C++17 standard to largely improve the expressivity and efficiency of the codebase.

Download

Taskflow 3.0.0 can be downloaded from here.

System Requirements

To use Taskflow v3.0.0, you need a compiler that supports C++17:

• GNU C++ Compiler at least v7.0 with -std=c++17
• Clang C++ Compiler at least v6.0 with -std=c++17
• Microsoft Visual Studio at least v19.27 with /std:c++17
• AppleClang Xode Version at least v12.0 with -std=c++17
• Nvidia CUDA Toolkit and Compiler (nvcc) at least v11.1 with -std=c++17
• Intel C++ Compiler at least v19.0.1 with -std=c++17

Taskflow works on Linux, Windows, and Mac OS X.

Working Items

• enhancing the taskflow profiler (TFProf)
• adding methods for updating tf::cudaFlow (with unit tests)
• adding support for cuBLAS
• adding support for cuDNN
• adding support for SYCL (ComputeCpp and DPC++)

New Features

Taskflow Core

• replaced all non-standard libraries with C++17 STL (e.g., std::optional, std::variant)
• added tf::WorkerView for users to observe the running works of tasks
• added asynchronous tasking (see Asynchronous Tasking)
• modified tf::ObserverInterface::on_entry and tf::ObserverInterface::on_exit to take tf::WorkerView
• added a custom graph interface to support dynamic polymorphism for tf::cudaGraph
• supported separate compilations between Taskflow and CUDA (see Compile Taskflow with CUDA)
• added tf::Semaphore and tf::CriticalSection to limit the maximum concurrency
• added tf::Future to support cancellation of submitted tasks (see Request Cancellation)

cudaFlow

• added tf::cudaFlowCapturer for building a cudaFlow through stream capture (see GPU Tasking (cudaFlowCapturer))
• added tf::cudaFlowCapturerBase for creating custom capturers
• added tf::cudaFlow::capture for capturing a cudaFlow within a parent cudaFlow
• added tf::Taskflow::emplace_on to place a cudaFlow on a GPU
• added tf::cudaFlow::dump and tf::cudaFlowCapturer::dump to visualize cudaFlow
• added tf::cudaFlow::offload and update methods to run and update a cudaFlow explicitly
• supported standalone cudaFlow (see Use cudaFlow in a Standalone Environment)
• supported standalone cudaFlowCapturer (see Use cudaFlow Capturer in a Standalone Environment)
• added tf::cublasFlowCapturer to support cuBLAS (see LinearAlgebracublasFlowCapturer)

Utilities

• added utility functions to grab the cuda device properties (see cuda_device.hpp)
• added utility functions to control cuda memory (see cuda_memory.hpp)
• added utility functions for common mathematics operations
• added serializer and deserializer libraries to support tfprof
• added per-thread pool for CUDA streams to improve performance

Taskflow Profiler (TFProf)

• added visualization for asynchronous tasks
• added server-based profiler to support large profiling data (see Profile Taskflow Programs)

New Algorithms

CPU Algorithms

• added parallel sort (see Parallel Sort)

GPU Algorithms

• added single task (see Single Task)
• added parallel iterations (see Parallel Iterations)
• added parallel transforms (see Parallel Transforms)
• added parallel reduction (see Parallel Reduction)

Bug Fixes

• fixed the bug in stream capturing (need to use ThreadLocal mode)
• fixed the bug in reporting wrong worker ids when compiling a shared library due to the use of thread_local (now with C++17 inline variable)

Breaking Changes

• changed the returned values of asynchronous tasks to be std::optional in order to support cancellation (see Asynchronous Tasking and Request Cancellation)

Deprecated and Removed Items

• removed tf::cudaFlow::device; users may call tf::Taskflow::emplace_on to associate a cudaflow with a GPU device
• removed tf::cudaFlow::join, use tf::cudaFlow::offload instead
• removed the legacy tf::Framework
• removed external mutable use of tf::TaskView

Documentation

• added Compile Taskflow with CUDA
• added Benchmark Taskflow
• added Limit the Maximum Concurrency
• added Asynchronous Tasking
• added GPU Tasking (cudaFlowCapturer)
• added Request Cancellation
• added Profile Taskflow Programs
• added cudaFlow Algorithms
  • Single Task to run a kernel function in just a single thread
  • Parallel Iterations to perform parallel iterations over a range of items
  • Parallel Transforms to perform parallel transforms over a range of items
  • Parallel Reduction to perform parallel reduction over a range of items
• added Governance
  • Rules
  • Team
  • Code of Conduct
• added Contributing
  • Guidelines
  • Contributors
• revised Conditional Tasking
• revised documentation pages for files

Miscellaneous Items

We have presented Taskflow in the following C++ venues with recorded videos:

• 2020 CppCon Taskflow Talk
• 2020 MUC++ Taskflow Talk

We have published Taskflow in the following conferences and journals:

• Tsung-Wei Huang, "A General-purpose Parallel and Heterogeneous Task Programming System for VLSI CAD," IEEE/ACM International Conference on Computer-aided Design (ICCAD), CA, 2020
• Chun-Xun Lin, Tsung-Wei Huang, and Martin Wong, "An Efficient Work-Stealing Scheduler for Task Dependency Graph," IEEE International Conference on Parallel and Distributed Systems (ICPADS), Hong Kong, 2020
• Tsung-Wei Huang, Dian-Lun Lin, Yibo Lin, and Chun-Xun Lin, "Cpp-Taskflow: A General-purpose Parallel Task Programming System at Scale," IEEE Transactions on Computer-aided Design of Integrated Circuits and Systems (TCAD), to appear, 2020