ME964High Performance Computing for Engineering ApplicationsWe live in a society exquisitely dependent on science and technology, in which hardly anyone knows anything about science and technology. Carl Sagan© Dan Negrut, 2011ME964 UW-MadisonBuilding CUDA apps under Visual Studio Accessing NewtonCUDA Programming ModelCUDA APIFebruary 03, 2011Before We Get Started… Last time HPC vs. HTC Quick overview of the CPU/GPU hardware and latency/bandwidths of relevance GPGPU, GPU programming and CUDA Started discussion on building CUDA apps in Visual Studio 2008 Today Andrew: wrap up building CUDA apps in Visual Studio 2008 Andrew: running apps through the HPC scheduler on Newton End quick overview of CUDA programming model Start discussion of CUDA API HW Due next Tu: two problems Includes reading of two papers: Amdhal and Manferdelli (links on the class website)23Setting up a Visual Studio 2008 CUDA ProjectBefore We Get Started… Assumptions Visual Studio 2008 installed CUDA Toolkit and GPU Computing SDK 3.2 installed (Optional) Developer Drivers installed Overview Set system environment variables for missing DLLs Run CUDA VS Wizard Set up project properties Compile and run example CUDA program4Environment Setup Why? Some DLLs missing from the %PATH% Add ;%NVSDKCOMPUTE_ROOT%\C\common\lib to system PATH environment variable Under: My Computer ->System Properties ->Advanced ->Environment Variables Alternatively: copy any missing DLLs to output dir5Configuring the Project Install CUDA VS Wizard* (http://goo.gl/Fh55o) Start Visual Studio, File -> New Project CUDA{32|64} -> CUDAWinApp, give name Next -> select Empty Project -> Finish Right click project -> Properties (if using shrUtils from Nvidia) Linker -> General Add ‘$(NVSDKCOMPUTE_ROOT)\shared\lib’ to Add’l Lib Dirs Linker -> Input Add shrUtils{32|64}{|D}.lib to Add’l Deps (Choose 32 or 64, nothing or D if release or debug) CUDA Build Rule -> General Add ‘$(NVSDKCOMPUTE_ROOT)\shared\inc’ to Add’l Include Dirs6* (Preferable) Use CUDA build rules directly, but the Wizard sets various other options for youWriting Code F7 to compile, F5 to debug (if you have a CUDA-capable GPU) Once it works, copy to Newton (next)7Helpful Hints(Compiling) Missing header (*.h) files: Add path to the ‘include’ dir containing the header Symbol not found Add *.lib file under dependencies, may need to add library dir(When running program) *.dll not found Either add the DLL dir to system path or copy DLL to program dir Black (cmd) window opens, quickly disappears Run from command prompt or (dangerous if on cluster) add a system(“pause”) at the end8Helpful Hints (cont’d) For code completion (‘Intellisense’) and pretty colors: Run *.reg file in %CUDA_PATH%\extras\visual_studio_integration In Visual Studio, Tools -> Options -> Text Editor -> File Extension Extension: cu Editor: Microsoft Visual C++910Running a CUDA app on the GPU cluster (Newton)Some Quick Notes… Must be inside the College of Engineering firewall to access Newton From a CAE lab On UWNet-Engineering wireless Via Engineering VPN (not WiscVPN) Remote Desktop via Kaviza - http://remote.cae.wisc.edu/dt/ User accounts managed by CAE All students registered for the class are eligible Auditors from outside CoE can request a temporary account Will see ‘ME964’ under your groups in My.CAE once you have access to the cluster This presentation will cover how to submit jobs via Remote Desktop Future presentation will show how to install HPC Pack and submit jobs from your local machine11Getting Started Copy all files for your program to Newton \\newton.msvc.wisc.edu\Data\ME964\%username% (replace %username% with your CAE username)12Remote Desktop to Newton Computer: newton.msvc.wisc.edu User name: ENGR\%username%13Start HPC Job Manager Start -> All Programs -> Microsoft HPC Pack 2008 R214Creating a Job New Single-Task Job Command line: program to run Working directory: \\newton\Data\ME964\%username% Give a filename for STDERR & STDOUT, though STDIN is optional Select number of cores to use Your program must be written to take advantage of them (egvia threads)15Creating a Job - Notes Single-Task Job simplest to setup New Job gives you much more control Parametric Sweep lets you run the same program with different parameters (Monte Carlo) GPUs are not currently reserved – be careful Working on this, HPC Pack does not natively let you do it All files for your program must reside in the working directory – unlike Condor, HPC Pack does not take care of this for you16Finishing Up After submitting, you can monitor the job’s progress in the Job Manager Once it finishes (or fails), double click for more info Which tasks finished/failed Task outputs Where each task ran17Why Did It Fail? Most common: libraries not installed 2ndmost common: compiled using wrong version of CUDA Toolkit When in doubt, reinstall from \\newton.msvc.wisc.edu\Data\Downloads\Drivers Does it run locally? Check log files for STDOUT and STDERR Still not working? Contact Andrew18Other Notes Programs must be able to finish/die without any user interaction, otherwise will hang OpenGL/DirectX are not available TCC Mode enabled, cards are only good for number crunching Also see: http://technet.microsoft.com/en-us/library/ff919670(WS.10).aspx1920Back to the Overview of the CUDA Programming ModelA Simple C-CUDA Program You want to add two vectors A and B and store the result in a vector C Assume that the size of the vectors is N=512 Here’s how things get done (some details omitted such as #define N 512) 21Execution Configuration: Grids and Blocks A kernel is executed as a grid of blocks of threads All threads in a kernel can access several device data memory spaces A block [of threads] is a batch of threads that can cooperate with each other by: Synchronizing their execution Efficiently sharing data through a low latency shared memory Threads from two different blocks cannot cooperate!!! This has important software design implicationsHostKernel 1Kernel 2DeviceGrid 1Block(0, 0)Block(1, 0)Block(2, 0)Block(0, 1)Block(1, 1)Block(2, 1)Grid 2Block (1, 1)Thread(0, 1)Thread(1, 1)Thread(2, 1)Thread(3,