Download presentation
Presentation is loading. Please wait.
Published byJeffry Ferguson Modified over 9 years ago
1
Parallelizing Sobel Edge Detection Message-Passing, Shared-Memory, and Streaming Implementations Patrick Griffin, Levente Jakob, and James Psota
2
Methodology 3 Computational Models: Message Passing Shared Memory Stream Evaluated on Raw simulator
3
Edges are defined by first derivative Find dx and dy with convolution Edge strength: We used: Sobel Edge Detection G y =G x =
4
Message Passing Implementation Calc Scatter + Calc Calc Gather + Calc Calc DRAM Tile8 distributes “oversized” partitions
5
Message Passing Results 177 cycles / pixel Memory copies kill performance Packaging/unpackaging messages expensive Alternative solution: send many messages, no need to package rMPI status Seemingly correct, but currently not optimized Non-blocking receives will help tremendously for some applications
6
Shared Memory Implementation
7
Shared Memory Results
8
Streaming Implementation Red Row Sum GxGx GyGy Green Row Sum GxGx GyGy Blue Row Sum GxGx GyGy PackSumAbs
9
Streaming Results 20 cycles / pixel Throughput limited by G x calculations Most nodes required no memory, G x and G y need 2 rows worth Integer ops + abs(float) All routing done on static network 1
10
Conclusions Streaming 4x slower than ideal Overhead of MP and SM detrimental MethodCycles / Pixel Ideal79/16 = 5 Streaming20 Shared Memory 539 Message Passing 177
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.