20131107
Download as PPTX, PDF0 likes218 views
This weekly report discusses a new method for computing integral images using CUDA that improves performance. Specifically, it processes one pixel per thread using shared memory to store intermediate values from the previous row, calculating the current row and storing the result back to shared memory to process the next row. It also describes using Nsight, an IDE for CUDA applications, to debug issues when problems arise with the new compute by row method.
![NEW METHOD (CONT.)
• Shared memory: store the values of the previous line
• computing by Row for img[*][y] and img[*][y+1]
• Time t: calculate img[*][y] + shared memory[*]
• Then store the result back to shared memory[*]
• Time t+1: calculate img[*][y+1] + shared memory[*]](https://image.slidesharecdn.com/20131107tpy-131115021349-phpapp01/85/20131107-4-320.jpg)
20131107
- 1. WEEKLY REPORT Thur., Nov 7, 2013 Pin Yi Tsai
- 2. OUTLINE • Current Work • New Method • Compute Integral Image – computeByRow unknown problem • How to Debug? Nsight
- 3. NEW METHOD • Previous Version: • One line per thread • Take too much time on accessing inconsistent memory • New one: • One pixel per thread • Using shared memory to store the middle value
- 4. NEW METHOD (CONT.) • Shared memory: store the values of the previous line • computing by Row for img[*][y] and img[*][y+1] • Time t: calculate img[*][y] + shared memory[*] • Then store the result back to shared memory[*] • Time t+1: calculate img[*][y+1] + shared memory[*]
- 5. COMPUTE INTEGRAL IMAGE – COMPUTE_BY_ROW • Original Input
- 6. COMPUTE INTEGRAL IMAGE (CONT.) • Correct Output – Compute By Row
- 7. COMPUTE INTEGRAL IMAGE (CONT.) • With problems
- 8. HOW TO DEBUG ? • Nsight ─ an IDE for developers building CUDA application ─ able to Profile and observe the value of each thread • In process: • Compile with custom makefile
- 9. HOW TO DEBUG ? (CONT.)
- 10. The End