|
hyPACK-2013 : Prog. on HPC GPU Cluster : Pthreads & OpenCL Prog.
|
|
High Performance computing GPU Cluster (HPC GPU Cluster) is based on integration of different programming paradigms on host-CPU
and device-GPU is available for laboratory sessions.
A prototype HPC GPU Cluster supports various heterogeneous programming features and itcan be made "adaptive" to the application it is running, assigning
the most effective resources in real-time as per application demands, without requiring modifications to the
application that is written using different progrmaming is used.
he Open Computing Language is a framework for writing programs that execute across heterogeneous platforms
consisting of CPUs, GPUs, and other processors.
Some of the following example programs use Pthread Programming environment on
host-CPU & and OpenCL on single/multiple GPUs in HPC GPU Cluster.
|
|
Example 1.1
|
To write a Pthreads - OpenCL program to
opencl-pthread program to perform Vector-Vector addtition that executes on multiple GPUs.
|
|
Example 1.2
|
To write a Pthreads - OpenCL program to
opencl-pthread program to perform Matrix-Vector Multiplication that executes on multiple GPUs.
|
|
The HPC GPU Cluster aim is
to develop system software and integrate components of the OpenCL on GPUs
and demonstrare some of the programs in Mode-2 & Mode-3. Different programming paradigms on host CPU i.e. MPI, OpenMP, Pthreads and OpenCL enabled
NVIDIA/AMD-APP are integrated and several
example programs for Matrix Computations, and Solution of Partial Differential
equations are discussed.
|
|
Description of Pthreads - OpenCL Programs |
- Objective
Write a Pthreads OpenCL program for Vector Vector multiplication on
multiple GPUs
- Description
The input data is paritioned on p threads and each P
each hread work is transfered from host-cPU to a device-GPU. Each thread
performs vector into vector addition on each GPU by called OpencL Kernel
and writes the partial sum to a result vector.
-
Implementation of Vector Vector Multiplication :
Step 1 : The memory for Arrays VectorA , VectorB and
ResultVector are allocted on host-CPU. Two additional vectors
are used on host-CPU to store the partial result of Vector Vector Multiplication that is performed
by each thread.
Step 2 : Main thread initializes the arrays VectorA , VectorB and
ResultVector .
Fill the arrays VectorA and VectorB with real single precsion values and
initialize the array ResultVector.
Step 3 : ThreadPart which decides how many elements each thread is to access in the VectorA & VectorB
is computed.
Step 4 : Memory is allocated for MyResultVector by each thread on host-CPU .
Step 5 : Input VectorA and VectorB are filled with required number of single real precsion vlues for
each thread Id host-CPU
Step 6 : Checking for the availability of device is done on host-CPU
Step 7 : Each thread computes the vector vector addition by calling the OpenCL kernel
VectorVectorAddtition and copies back the result to MyResultVector on host-CPU .
Step 8 : Each thread performs a mutex operation on the result vector and assigns the
corresponding elements on host-CPU.
Step 9 : Main thread waits for all threads to exit and
Main thread prints the resultant vector.
-
Input
Master thread on host-cpu reads the input parameter n , the size of the vector
and the number of threads is p. The input vectors generated using random numbers
-
Output
Main thread prints the final vector
|
- Objective
Write a Pthreads OpenCL program for vector vector multiplication on
multiple GPUs
- Description
The input matrix is partioned on p threads and
each thread on host-CPU is transfered the partitioned matrix and vector
from host-cPU to a device-GPU. Each GPU
performs parttioned matrix into vector mutliplication corresponds to thrread Id
by calling OpencL Kernel and accumaltes partial result vector.
-
Implementation of Matrix Vector Multiplication :
Step 1 : Three Arrays MatrixA , VectorB and
ResultVector are declaed on host-CPU. Two additional vectors
are used on host-CPU for computation. The purpose is
to to store the part of the matrix that is accessible
by each thread and the other is to store part of the result vector computed by each thread.
Step 2 : Main thread initializes the arrays MatrixA, VectorB and
ResultVector .
Fill arrays MatrixA and VectorB with single precsion real values and
initialize the array ResultVector.
Step 3 : ThreadPart which decides how many elements each thread is to access in the matrix is computed.
Step 4 : Memory is allocated for MyMatrixA and MyResultVector by all the threads on host-CPU .
Step 5 : MyMatrixA is constructed from MatrixA depending on the value of thread Id host-CPU
Step6 : Memor for Vectors are allocated on host-CPU . The values of the vectors
in the host machine are copied to the vectors on the device machine.
Step 7 :Checking for the availability of device is done on host-CPU
Step 8 : Each thread computes the Matrix vector multiplication by calling the CUDA kernel
MatrixVectorMultiplication and copies back the result to MyResultVector on host-CPU .
Step 9 : Each thread performs a mutex operation on the result vector and assigns the corresponding elements
on host-CPU.
Step 10 : Main thread waits for all threads to exit and
Main thread prints the resultant vector.
-
Input
Master thread on host-cpu reads the input parameter n , the size of the vector
and the number of threads is p.
-
Output
Main thread prints the final vector
|
|
|
|
|
