Upload
todd-tyler
View
220
Download
0
Embed Size (px)
Citation preview
Applying GPU and POSIX Thread Technologies in Massive Remote Sensing Image Data Processing
By: Group 17 King Mongkut's Institute of Technology Ladkrabang
High Resolution graphics not Smoother?
1. Buy better Graphics Card
2. Install newest Driver
<< This Paper >>3. Better GPU & CPU Management!!
Why
This Paper?
GPU?POSIX?
- IEEE Thread Standard- Thread API for Many OS
- Pthread
= GPU + Thread Technique
Apply GPU & POSIX Thread
What
Overview
1.Solve?2.Technique3.Process4.Real App
1.Solve?
1.Solve?Problem: Remote sensing is Massive!
High Resolution = Slow Processing = Can’t Real Time
Goal: Processing image faster
2.Technique
2.Technique (1/4)
1. CUDA
2. Block and Tile
3. Dual-parallel Processing
2.Technique (2/4)1.CUDA
(Compute Unified Device Architecture)
Is: Parallel Architecture for using GPU
Use by: NVIDIA GPUs since 2007 e.g. GeForce GT 420*
2.Technique (3/4)2. Block and Tile
What is Block and Tile
● Image data share to block of I/O
● If block larger than gpu , have problem
2.Technique (4/4)3. Dual-parallel Processing
POSIX threads are applied to perform the I/O step and the GPU processing step simultaneously.
3.Process
2.Process(1/2)Main
Init of CUDA
Two Thread are set Up
Process of I/O Thread
And GPU Process
pthread_join
Return
3.Process (2/2)
4.Real App
EXPERIMENT
Database PostgreSQL
4.Real App
CPU 1 ( AMD Athon 3000+ , 1.81GHz ) CPU 2 ( Intel Core i3 530 , 2.93 GHz ) GPU ( Nvidia Geforce GT240 , 512M )
1 2 3
EXPERIMENT
4.Real App1 2 3
4
56
Image1
Image2
Image3
Image4
Origital Image
ProcessesProcessed Image
Input
Process
Output
4.Real AppOriginal image Processed image
4.Real App
IMAGE 1
IMAGE 2
IMAGE 4
IMAGE 3
4.Real AppImage123821 * 21758
Image212000 * 10961
Image36000 * 5480
Image43000 * 2740
- PostgreSQL is a powerful- Object-Relational database system- More Than MySQL
- CUDA technology directly to process the image blocks stored in database rapidly with SQL statements
4.Real AppDatabase PostgreSQL
4.Real App
1.Use2.Download3.Install
Database PostgreSQL
4.Real App1.Use PostgreSQL
PostPIC PostGIS
SELECT *FROM images WHERE where date(the_img) > '2009-01-01'::date and size(the_img) > 1600;
- Support geographic objects to PostgreSQL
- Simple Features Specification for SQL
- Types and Functions
- Download : http://www.postgresql.org/download
4.Real App2. Download
4.Real App3.1 Install PostgreSQL
1 2 3
4 5 6
4.Real App3.1 Install PostgreSQL
7 8 9
10 11 12
4.Real App 3.2 Install PostGIS
16 17 18
13 14 15
4.Real App3.2 Install PostGIS
19 20
4.Real App
PostgreSQL 9.3
PostGIS 2.1
3.3 Install Complete
Conclusion!
Conclusion1.Solve?= Processing
faster
3.Process= Flowchart
2.Technique= Dual-Process
4.Real App= Speed up
PostPIC PostGIS
PostgreSQL
Conclusion
CPU GPU
Block and Tile
Applying GPU and POSIX Thread Technologies in Massive Remote Sensing Image Data Processing
By: Group 17By: Group 17
THANK YOU !!
Applying GPU and POSIX Thread Technologies in Massive Remote Sensing Image Data Processing
Extended!!
1. Technique (Extended)
2. Process (Extended)
1. Technique X
1. Technique X
1.CUDA X
2.Block and Tile X
3.Dual-parallel Processing X
1. Technique X1.CUDA (Extended)
1. Technique X
1.1 Architecture
Confix GPU
API
1.CUDA (Extended)
1. Technique X
1. Language integration-Level = e.g. C, C++, …
2. Device-Level = e.g. Assembly
1.1 Architecture1.CUDA (Extended)
-> API
1. Technique X
1.2 Parallel
CPU GPU
1.CUDA (Extended)
1. Technique X
- A resource management program in OS
-> CPU
- Allow copy backhelp other compute
1.2 Parallel
1.CUDA (Extended)
1. Technique X
Partition of imageData in GPU
Partition for a kernel
Processing data in a Block
-> GPU1.2 Parallel
1.CUDA (Extended)
1. Technique X2. Block and Tile
(Extended)
1. Technique X2. Block and Tile
(Extended)
• The block is the I/O unit between sensing image data and the system memory
1. Technique XBlock and Tile(Cont)
• We tested the I/O performance with different block sizes
• Results show that the I/O performance of sensing image data declines
• The tile is used to transfer data between the system memory and GPU memory
• Block , the I/O unit between image and memory
• Tile , the I/O unit between memory and GPU memory
1. Technique XBlock and Tile(Cont)
• The block in system memory is partitioned into multiple tiles in this approach
• The tile is used to transfer data between the system memory and GPU memory
1. Technique X2. Dual-parallel
Processing (Extended)
3. Dual-parallel Processing (Extended)
1. Technique X
• Using buffer pool technology.
• I/O step and the GPU processing step are independent from each other.
• Simple and easy to implement if buffer size is equal to block size.
1.Technique X
• Used for the I/O task between the image data and the system memory
• Responsible for delivering the buffers from the buffer pool to the GPU memory and processing
3. Dual-parallel Processing (Extended)
1.Technique X3. Dual-parallel Processing (Extended)
1.Technique X
• From the micro perspective, image data are processed by hundreds of execution units simultaneously in GPU
• From the macro perspective, the I/O operation and the processing operation are performed simultaneously by POSIX threads.
3. Dual-parallel Processing (Extended)
1.Technique X3. Dual-parallel Processing (Extended)
2.Process X
2.Process(1/3)• Begins with function main after that Initialization of CUDA
environment• Group of buffers are set up• Two POSIX threads are created• One is used to input and output the remote sensing image data• Two is used to responsible for processing the buffers read by
the first thread• For example
• Ready_to_read• The I/O thread should read a block from image data
to current buffer• Ready_to_process
• Buffer is ready to be process by GPU• Ready_to_process
• The I/O thread should write a block to image data
2.Process(2/3)• Two global variables are declared to record the execution state
of each thread• For example
• Is_IO_Over• is true, it means all the work of I/O thread is finished
• Is_Process_Over• is true, it means all the work of responsible for
processing the buffers is finished• “pthread_join” would be called to terminate that thread• Two threads are finished then all the remote sensing image
data are completely processed and the program will be end by calling “return”