Difference between revisions of "Multi-threaded Programming"
From Gridkaschool
ThomasHauth (talk | contribs) |
ThomasHauth (talk | contribs) |
||
| Line 1: | Line 1: | ||
= Introduction = |
= Introduction = |
||
| + | Intel Threading Building Blocks (TBB) is an open-source C++ library to support |
||
| − | OpenCL is a standard which defines a framework, an API and a programming language for parallel computation on heterogeneous systems like client computer systems, high- performance computing servers as well as hand-held devices. |
||
| + | the development of mulit-threaded applications which can exploit the available |
||
| + | processing cores in modern CPUs. Therefore, various contsructs to support |
||
| + | parallelism in applications are provided by TBB. Low-level constructs are |
||
| + | available to partition loops to run them distributed over many CPU cores. |
||
| + | High-level constructs like a task scheduler and a graph-based execution model |
||
| + | allow to express dependencies and relations between computing tasks and TBB can |
||
| + | distribute these items among the available cores. |
||
| + | Furthermore, C++ classes are included that guarantee a thread-safe access to |
||
| − | The standard is maintained by the Khronos Group and supported by a large consortium of industry leaders including Apple, Intel, AMD, NVIDIA and ARM. Influenced by NVIDIA’s CUDA from the GPU side and by OpenMP which originates from the classical CPU side, the open OpenCL standard is characterized by a formulation which is abstract enough to support both CPU and GPU computing resources. |
||
| + | containers like lists and maps. Also explicit locking constructs like mutexes |
||
| + | are available. |
||
| + | The TBB library supports the Windows, Mac OS and Linux operating systems and the |
||
| − | This is an ambitious goal, since providing an abstract interface together with a peak performance is a challenging task. OpenCL employs a strict isolation of the computation work into fundamental units, the kernels. These kernels can be developed in the OpenCL C programming language, a subset of the C99 language, with some additional OpenCL specific keywords. |
||
| + | Visual C++, GCC and Intel compilers. |
||
| − | |||
| − | In general, these kernels are hardware independent and compiled by the OpenCL runtime when they are loaded. To be able to fully exploit the parallel execution of the kernel code, several kernel instances, the work items, are started to process a set of input values. The actual number of concurrently running work items is determined by the OpenCL system. How a concrete algorithm can be partitioned into work items has to be decided by the programmer. |
||
= Reference Material = |
= Reference Material = |
||
<ul> |
<ul> |
||
| − | <li><p> |
+ | <li><p>Intel Threading Building Blocks - Tutorial</p> |
| − | <p>http:// |
+ | <p>http://threadingbuildingblocks.org/uploads/81/91/Latest\%20Open\%20Source\%20Documentation/Tutorial.pdf</p></li> |
| − | <li><p> |
+ | <li><p>Intel Threading Building Blocks - Reference</p> |
| − | <p>http:// |
+ | <p>http://threadingbuildingblocks.org/uploads/81/91/Latest\%20Open\%20Source\%20Documentation/Reference.pdf</p></li> |
| − | <li><p> |
+ | <li><p>Intel Threading Building Blocks - Design Patterns</p> |
| − | <p>http:// |
+ | <p>http://threadingbuildingblocks.org/uploads/81/91/Latest\%20Open\%20Source\%20Documentation/Design_Patterns.pdf</p></li> |
| + | </ul> |
||
| − | <li><p>Intel SDK for OpenCL</p> |
||
| − | <p>http://software.intel.com/en-us/articles/vcsource-tools-opencl-sdk/</p></li> |
||
| − | <li><p>AMD OpenCL Zone</p> |
||
| − | <p>http://developer.amd.com/zones/OpenCLZone/</p></li> |
||
| − | <li><p>NVIDIA OpenCL</p> |
||
| − | <p>http://www.nvidia.com/object/cuda_opencl_1.html</p></li></ul> |
||
| − | |||
| − | = Programming Environment and Libraries = |
||
| − | |||
| − | c++ interface / blah |
||
| − | |||
| − | = Project: Boostraping OpenCL and Vector Addition = |
||
| − | |||
| − | TODO: give overview of what to do |
||
| − | |||
| − | == Compiling and running the test program == |
||
| − | |||
| − | Open the folder '''project_vectoradd''', create the build files using CMake and compile the application. |
||
| − | |||
| − | |||
| − | |||
| − | <pre>$ cd project_vectoradd/ |
||
| − | [hauth@vdt-corei7avx project_vectoradd]$ cmake . |
||
| − | [hauth@vdt-corei7avx project_vectoradd]$ make |
||
| − | [100%] Built target vectoradd |
||
| − | [hauth@vdt-corei7avx project_vectoradd]$ ./vectoradd </pre> |
||
| − | Once the application was compiled successully run it. The output should be along the following lines: |
||
| − | |||
| − | <pre>$ ./vectoradd |
||
| − | Testing Platform : Intel(R) OpenCL |
||
| − | > Selected Compute Device : Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz |
||
| − | Transferring data to device memory took 2e-06 s |
||
| − | Running vectorAdd kernel took 2e-06 s |
||
| − | Transferring data to host memory took 0 s |
||
| − | All done</pre> |
||
| − | == Task: Understand the existing code == |
||
| − | |||
| − | The first task is to read and understand the existing source code. To do so, you have to edit the file <tt>vectoradd.cpp</tt> in your favorite text editor, use <tt>nano</tt> if you are not sure which tool to use. |
||
| − | |||
| − | <pre>$ nano vectoradd.cpp</pre> |
||
| − | Take your time to familiarize yourself with the sourcecode which is already in the file. Some of the intial steps of setting up the OpenCL system are already provided: |
||
| − | |||
| − | <ul> |
||
| − | <li><p>OpenCL compute context</p> |
||
| − | <p>An OpenCL platform is automaticaly selected, depending on the required device type (CPU/GPU) and a compute context is created. You can change the required device type by modifying the constant <tt>devType</tt>:</p> |
||
| − | <pre>// Desired Device type. |
||
| − | // can be CL_DEVICE_TYPE_GPU or CL_DEVICE_TYPE_CPU in this example |
||
| − | const cl_device_type devType = CL_DEVICE_TYPE_CPU;</pre></li> |
||
| − | <li><p>OpenCL command queue</p> |
||
| − | <p>A command queue is created on the context</p></li> |
||
| − | <li><p>Kernel compile</p> |
||
| − | <p>The variable contains <tt>kernelSourceAdd</tt> the source code of the OpenCL kernel. This string ist used to compile and register the kernel with the OpenCL runtime system. If you want to extend the kernel code, you have to add to the string contained in the variable <tt>kernelSourceAdd</tt>.</p></li> |
||
| − | <li><p>Host data buffers</p> |
||
| − | <p>The variables <tt>hostVectorA</tt>, <tt>hostVectorB</tt> and <tt>hostVectorC</tt> hold buffers which will be used later to transfer data to and from the OpenCL kernel. The underlying C++ type is <tt>std::vector</tt> which can be conviniently used to acquire a contingious space of memory in host’s RAM.</p> |
||
| − | <pre>FloatingPoint vp = 23.0f; |
||
| − | FloatingPointVector hostVectorA(itemCount, vp); |
||
| − | FloatingPointVector hostVectorB(itemCount, vp); |
||
| − | FloatingPointVector hostVectorC(itemCount, vp);</pre></li> |
||
| − | <li><p>Various timing measurements</p> |
||
| − | <p>To quantify the runtime of the OpenCL operations, various timing objects are present. For the time measurements to be correct, only the operations indicated by the TODO comments must be between the creation of the <tt>Timing</tt> object and the call to <tt>EndWithReport()</tt> of the respective timing object.</p> |
||
| − | <pre>Timing t_transfer_input("Transferring data to device memory"); |
||
| − | // TODO: transfer input buffer here |
||
| − | t_transfer_input.EndWithReport();</pre></li></ul> |
||
| − | |||
| − | == Task: Running the OpenCL Kernel == |
||
| − | |||
| − | The OpenCL kernel is already created and associated with the compiled program: |
||
| − | |||
| − | <pre>cl::Kernel kernelAdd(pProgram, "vectorAdd");</pre> |
||
| − | Use the <tt>enqueueNDRangeKernel(…)</tt> method on the command queue object to add a kernel instance to the queue <ref>http://www.khronos.org/registry/cl/sdk/1.2/docs/man/xhtml/clEnqueueNDRangeKernel.html |
||
| − | </ref> |
||
| − | |||
| − | Call the <tt>finish()</tt> method on the command queue object to ensure the kernel is done processing before the host program continues execution. <ref>http://www.khronos.org/registry/cl/sdk/1.2/docs/man/xhtml/clFinish.html |
||
| − | </ref> |
||
| − | |||
| − | == Modifications to play around == |
||
| − | |||
| − | * switch to double , how does the runtime change for CPU/GPU ? |
||
| − | * switch to float4 vector type. can you perform the same addition operations ? |
||
| − | |||
| − | = Project: N-Body Simulation = |
||
| − | |||
| − | <references /> |
||
| − | |||
| − | |||
| − | ===[[Internals:Multi-threaded|Technical specification/requirements]]=== |
||
Revision as of 04:26, 24 August 2012
Introduction
Intel Threading Building Blocks (TBB) is an open-source C++ library to support the development of mulit-threaded applications which can exploit the available processing cores in modern CPUs. Therefore, various contsructs to support parallelism in applications are provided by TBB. Low-level constructs are available to partition loops to run them distributed over many CPU cores. High-level constructs like a task scheduler and a graph-based execution model allow to express dependencies and relations between computing tasks and TBB can distribute these items among the available cores.
Furthermore, C++ classes are included that guarantee a thread-safe access to containers like lists and maps. Also explicit locking constructs like mutexes are available.
The TBB library supports the Windows, Mac OS and Linux operating systems and the Visual C++, GCC and Intel compilers.
Reference Material
Intel Threading Building Blocks - Tutorial
Intel Threading Building Blocks - Reference
Intel Threading Building Blocks - Design Patterns
