CUDA Development using

PyCUDAPart 1

Prof. Mario A. Gazziro (Yah!)Organization: Prof. André

Carvalho

July 2012Support:

Igor, Heitor, Pedro, Ruan and Andre

What 1 TeraFlop/s means ?

Computer TeraFlops/s

Year of instalatio

n

PriceUS$

Institute

PowerPC970

3 2007 500.000 USP (CCE)

Cluster GPU

Attilio *

16 2009 100.000 IFSC

BlueGene L 26 -(49) 2007-(2010)

Donation IINN

TUPÃ 258 2010 25.000.000 INPE

*only paralell computer – all other are serial ones.

Instalation – Step 1DRIVER:sudo /etc/init.d/gdm stop<ALT+F1><logar como labredes senha 12345678>chmod 777 devdriver_4.2_linux_64_295.41.runsudo ./devdriver_4.2_linux_64_295.41.run<concordar com tudo>sudo /etc/init.d/gdm start

Instalation – Step 2TOOLKIT:chmod 777 ./cudatoolkit_4.2.9_linux_64_ubuntu10.04.runsudo ./cudatoolkit_4.2.9_linux_64_ubuntu10.04.run <concordar com todas as opcoes><incluir o texto abaixo no final do arquivo .bashrc>cd ~gedit .bashrcexport PATH=/usr/local/cuda/bin:$PATHexport LPATH=/usr/lib/nvidia-current:$LPATHexport LIBRARY_PATH=/usr/lib/nvidia-current:$LIBRARY_PATHexport

LD_LIBRARY_PATH=/usr/lib/nvidia-current:/usr/local/cuda/lib64:/usr/local/cuda/lib: $LD_LIBRARY_PATH

<salvar e sair>

Instalation – Step 3SDK:chmod 777 gpucomputingsdk_4.2.9_linux.run<NAO USAR SUDO!>./gpucomputingsdk_4.2.9_linux.run<concordar com todas as opcoes><fechar todas janelas de terminal - variaveis de ambiente>

<Testar compilador>nvcc<deve aparecer a seguinte mensagem de erro>nvcc fatal : No input files specified; use option --help

for more information

Instalation – Step 4cd ~/NVIDIA_GPU_Computing_SDK/make<aguardar>cd ~/NVIDIA_GPU_Computing_SDK/C/bin/linux/release/ls./SobelFilter./particles<clique com botao 2 e selecione move cursor mode>< mova o mouse>./fluidsGL<FECHAR TODOS>./particles<CTRL+Z>bg./fluidsGL< verificar execução conconcorrente dos kernels >< compartilhamento dos recursos da GPU >

CUDA SDK sample applications

CUDA SDK sample applications

Instalation – Step 5<INSTALACAO DAS DEPENDENCIAS DO PYTHON >sudo apt-get install python-numpysudo apt-get install python-h5py sudo apt-get install python-scipysudo apt-get install python-matplotlib<testar interface grafica>cd ~/Área de Trabalho/CUDApython ./teste_grafico.py

visualizador hdf5:chmod 777 hdfview_install_linux64.bin./hdfview_install_xxx.bin./hdfview

Instalation – Step 6Py-Cuda:sudo apt-get install build-essential python-dev python-

setuptools libboost-python-dev libboost-thread-dev –ytar xzvf pycuda-2011.2.2.tar.gz cd pycuda-2011.2.2/./configure.py --cuda-root=/usr/local/cuda --cudadrv-lib-

dir=/usr/lib --boost-inc-dir=/usr/include --boost-lib-dir=/usr/lib --boost-python-libname=boost_python-mt --boost-thread-libname=boost_thread-mt --no-use-shipped-boost

make -j 4<apagar siteconf.py em caso de erro!!!>sudo env PATH=$PATH python setup.py install<testar pycuda>cd ~/Área de Trabalho/CUDA/pycuda-2011.2.2/examplespython ./demo.py<visualizar manipulacao de arrays e matrizes>

Part I: Overview

Definition:

Graphical Processing Units are just graphical card adaptors that can give access to programmers to their internal API (Advanced Program Interface). Today, there are even GPUs without graphics output (build only to perform scientific calculations).

Introduced in 2006, the Compute Unified Device Architecture is a combination of software and hardware architecture (available for NVIDIA G80 GPUs and above) which enables data-parallel general purpose computing on the graphics hardware. It therefore offers a C-like programming API with some language extensions.

Key Points:

The architecture offers support for massively multi threaded applications and provides support for inter-thread communication and memory access.

Why this topic is important?

Data-intensive problems challenge conventional computing architectures with demanding CPU, memory, and I/O requirements.

Emerging hardware technologies, like CUDA architecture can significantly boost performance of a wide range of applications by increasing compute cycles and bandwidth and reducing latency.

Where would I encounter this?

Gaming

Raytracing

3D Scanners

Computer Graphics

Number Crunching

Scientific Calculation

CUDA vs Intel

NVIDIA GeForce 8800 GTX vs Intel Xeon E5335 2GHz, L2 cache 8MB

Grid of thread blocks

The computational grid consist of a grid of thread blocks

Each thread executes the kernel

The application specifies the grid and block dimensions

The grid layouts can be 1, 2 or 3-dimensional

The maximal sizes are determined by GPU memory

Each block has a unique block ID

Each thread has a unique thread ID (within the block)

Elementwise Matrix Addition

Elementwise Matrix Addition

The nested for-loops are replaced with an implicit grid

Memory model

CUDA exposes all the different type of memory on GPU:

Part II: Classroom Exercises – HEAT (local)

Compile and run

heat.cu example

Command line: nvcc heat.cu –o heat -lglut

Part II: Classroom Exercises – HEAT (remote)

Compile and run heat.cu example in HAL9k gpu server (IFSC/CIERMag)

cmd: ssh – X yah@afrodite.ifsc.usp.br –p 2236

What is the change ?

Why this change happen ?

What is influence of the network latency in the final result ?

Part II: Classroom Exercises – EXERC1

Type, compile and test the following code. What this program do ?

Command line: gcc exerc1.c –o exerc1

Part II: Classroom Exercises – EXERC2

Type, compile and test the following code. What this program do ?

Command line: gcc exerc2.c –o exerc2

Part II: Classroom Exercises – EXERC3

Type, compile and test the following code. What this program do ?

Command line: nvcc exerc3.cu –o exerc3

Part II: Classroom Exercises – EXERC4

Part II: Classroom Exercises – EXERC5

Part II: Classroom Exercises – EXERC6 – part A

Part II: Classroom Exercises – EXERC6 – part B

Part III: Project

Case Study: Initial calculation for solving sparse matrix in the method proposed by professor Guilherme Sipahi, from IFSC

N=1001;K(1:N) = rand(1,N);g1(1:2*N) = rand(1,2*N);k = 1.3;tic;for i=1:N for j=1:N M(i,j) = g1(N+i-j)*(K(i)+k)*(K(j)+k); endend

Task: Design the CUDA kernel for this algorithm (using PyCuda or C) and compare its speed-up with the gold-standard provided by professor.

Part III: Project Best Solution (Mateus and Bié)

Case Study: Initial calculation for solving sparse matrix in the method proposed by professor Guilherme Sipahi, from IFSC

BLOCK(16, 2, 1)

GRID(1000/32, 1000)

370uS with 16 cuda cores

Questions ?

So long and thanks by all the fish! – See you tomorrow!

2nd day activities:

-Database integration

(HDF5)

-Graphics Visualization

(Matplotlib)

-Thread syncs and

Thread fences

- Atomic Operations and

Critical region control

References

Gokhale M. et al, Hardware Technologies for High-Performance Data-Intensive Computing, IEEE Computer, 18-9162, pg 60, 2008.

Lietsch S. et al. A CUDA-Supported Approach to Remote Rendering, Lecture Notes in Computer Science. 2007.

Fujimoto N. Faster Matrix-Vector Multiplication on GeForce 8800 GTX, IEEE, 2008.

Book Reference

NVIDIA Corporation, David, NVIDIA CUDA Programming Guide, Version 1.1, 2007.

/ s ?