FinalProject

Parallelize and optimize an application
4.0
fit@hcmus
Parallelizing/optimizing a program
What part(s) should be parallelized/optimized?
• Measure times of parts to decide
• When optimizing (after parallelizing), can measure times
of GPU activities quickly:
• nvprof --print-gpu-trace ./a.out
1. Analyze
chỗ nào cần song2 hóa • Each loop will create a
new version based on
previous versions
• We should go step by
song2 hóa như thế nào
step, from sequential to
parallel, from parallel to
4. Evaluate 2. Design optimized parallel
How to parallelize/optimize?
Does the idea work?
If not, do you know
why? 3. Implement
2
Parallelizing/optimizing a program
What part(s) should be parallelized/optimized?
• Measure times of parts to decide
• When optimizing (after parallelizing), can measure times
of GPU activities quickly:
• nvprof --print-gpu-trace ./a.out
How to go through this process

1. Analyze as well as possible?
Some advices:
• Keep the mind still
• Keep the code clean
• Code fast or slow?
4. Evaluate 2. Design • Use a good editor and learn
how to use it efficiently
How to parallelize/optimize?
Does the idea work?
If not, do you know
why? 3. Implement
3
General optimization guidelines
 Expose enough independent tasks to utilize
GPU hardware resources
 Expose enough blocks to utilize SMs
 In each SM, expose enough independent instructions
(coming from the same warp, or from different warps)
to utilize execution pipelines, hide latency
 Access DRAM efficiently
 Don’t let threads in the same warp access scattered
addresses in DRAM
 Use SMEM to reduce DRAM accesses, as well as to
access DRAM efficiently
 Reduce warp divergence
4
Final project - Contents of Colab notebook
1. Application description
 What is your chosen application?
đưa ví dụ vào (ảnh) input gì output là gì
 Input? Output?
 Use cases?
 Does it need to speed up?
5
2. Sequential implementation
 Design: Describe steps to go from input to output
(don’t show code) chi tiết nhất có thể
 Evaluate:
 Describe your experiment setup testing trên bộ gì, train/test bnh, sài GPU gì
 Run the code to see results

 Does it run correctly?
6
3. Parallel implementation naive version
 Analyze: Which steps do you parallelize? Why these

steps?
 Design: How do you parallelize? (don’t show code)
 Evaluate:
 Describe your experiment setup
 Does it run correctly & faster? If not, do you know why?
7
4. Parallel implementation + optimization

You should have ≥ 2 optimized versions
At each version:
 Analyze: Which parts (often: which kernels) do you
optimize? Why these parts?
 Design: How do you optimize? (don’t show code)
 Evaluate:
 Describe your experiment setup
 Does it run correctly & faster? If not, do you know why?
8
5. Reflection
 Each member: What difficulties have you encountered?
 Each member: What have you learned?
 Your team: If you had more time, what would you do?
9
6. References
To finish this project, what materials have you
consulted?
10
Final project - Code files
Each version (sequential version, parallel version,

1st optimized parallel version, 2nd optimized parallel
version, …) should be in a separate file
11
Final project - Teamwork
Your team should have a plan file
All members in your team should understand the

team’s project thoroughly (of course, it includes
code)
12
Final project - Submission & presentation
 x = presentation day
 x will be one day from xxx to xxx (I will decide and let you
know later)
 Before 23:55 day x-1: upload your team’s project to a link in
Moodle, include:
 Team plan file and work distribution
 Colab notebook file
 All source code file and an instruction file on how to set up and run
your project
 A presentation video about 15-20min. Upload on YouTube with
Unlisted option
 Day x: present offline in classroom (use Colab notebook file
to present, no need to prepare slides)
 Each team will have ~15 minutes to present (each member will
present ~1/2 contents, and I will decide who will present which)
and ~10 minute to Q & A
13
Thank you
14

FinalProject

Uploaded by

Copyright:

Available Formats

FinalProject

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

FinalProject

Uploaded by

Copyright:

Available Formats

Parallelize and optimize an application

How to go through this process

 Run the code to see results

3. Parallel implementation naive version

 Analyze: Which steps do you parallelize? Why these

4. Parallel implementation + optimization

Each version (sequential version, parallel version,

Your team should have a plan file

All members in your team should understand the

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.