Development Roadmap

[jensen-yan]

Here is the development roadmap. Contributions and feedback are welcome.

Project Roadmap

1. Dynamic Vector Length Support

• Implement variable vector length (not fixed at 128).

2. Value Prediction

• Loop iterators(Value prediction development roadmap #188 ).
• vl prediction.
• Generic value prediction(Value prediction development roadmap #188 ).

3. Speculative Memory Bypassing

• Table-less implementation.
• Dynamic bypassing.

4. Top-down Analysis Framework

• Further breakdown of details in top-down approach.
• Detailed documentation for top-down methodology.
• TIP/TEA integration.

5. Front-End Optimizations

• 8-wide instruction issue
• Two-taken branch predictor
• uBTB: Expand for improved timing guarantees.
• TAGE:
  • Increase from 5-table to 10-table cascaded comparison for better timing and reduced MPKI.
  • Compare increasing the number of levels versus increasing the number of entries.
• Fetch Pipeline:
  • Refactor for two-taken + two-fetch mechanism.
  • Enable single-cycle access to multiple cache lines.
  • Clean up fetch and cache pipeline code.

6. Calibration

• Detailed load/store pipeline: challenging, important, but not urgent. (Align lsq #186 )
• Calibrate functional unit (FU). (Align scheduler #185 )
• Calibrate scheduler. (Align scheduler #185 )
• Calibrate Memory speculation. (Developing)
• Calibrate Prefetchers. (Developing)

7.Hypervisor (H) Extension Improvements

• OpenSBI + kernel support.
• MMU & prefetch optimizations

8.Cliff - Microbenchmark for Calibration

• Cliff for frontend testing.
• Cliff for prefetcher evaluation.

How to Contribute to This Roadmap

1. To add a new task, use the following format:

   - [ ] Task description
   

2. To mark a task as completed, change [ ] to [x]:

   - [x] Completed task
   

3. To link to an issue or pull request, use the # symbol followed by the issue/PR number:

   - [ ] Implement feature X (#123)
   

• Implement feature X (A dummy pr to fix PR offset #123)

4. Feel free to add comments or additional information using HTML comment syntax:

   <!-- This is a comment that won't be visible in the rendered markdown -->
   

[zybzzz]

I am excited to see the RoadMap, and the toolchain provided by Xiangshan has significantly improved the execution speed of gem5 and shortened the debugging time.

I'm a current master's student in computer architecture, and my current work is focused on value prediction, but I'm also interested in top-down analysis methods. I'd be happy to contribute to xs-gem5, but considering that xs-gem5 is in the internal development phase, I am unsure whether it is open to contributions from ordinary developers in the open-source community.

If so, could I send an email to the appropriate people? I'll describe my background, what I can do, and what I'd like to do in the email.

@jensen-yan

[shinezyy]

I am excited to see the RoadMap, and the toolchain provided by Xiangshan has significantly improved the execution speed of gem5 and shortened the debugging time.

I'm a current master's student in computer architecture, and my current work is focused on value prediction, but I'm also interested in top-down analysis methods. I'd be happy to contribute to xs-gem5, but considering that xs-gem5 is in the internal development phase, I am unsure whether it is open to contributions from ordinary developers in the open-source community.

If so, could I send an email to the appropriate people? I'll describe my background, what I can do, and what I'd like to do in the email.

@jensen-yan

xs-gem5 has ended internal development phase, and is now open to contributions!

On contributing to topdown/value prediction, we can open another issue to discuss. And you can also e-mail me (zhouyaoyang@bosc.ac.cn or archshinez@outlook.com) and @jensen-yan

[muke101]

Hi, this is really cool and I hope a lot of progress can be made in implementing these more advanced techniques.

I've Implemented a newer MDP algorithm called PHAST (https://webs.um.es/aros/papers/pdfs/sskim-hpca24.pdf) based on another Gem5 fork. It achieves much higher accuracy than Store Sets (the open source version at least. I see this repo implements a better version that I'm keen to evaluate PHAST against too). It feels it'd be appropriate to aim to include the algorithm in this fork too. The repo is here: https://gitlab.com/muke101/gem5-phast

I wouldn't have time to integrate it to this repo myself, but I'd be happy to offer help to anyone who might want to take it on.

[happy-lx]

Hi, this is really cool and I hope a lot of progress can be made in implementing these more advanced techniques.

I've Implemented a newer MDP algorithm called PHAST (https://webs.um.es/aros/papers/pdfs/sskim-hpca24.pdf) based on another Gem5 fork. It achieves much higher accuracy than Store Sets (the open source version at least. I see this repo implements a better version that I'm keen to evaluate PHAST against too). It feels it'd be appropriate to aim to include the algorithm in this fork too. The repo is here: https://gitlab.com/muke101/gem5-phast

I wouldn't have time to integrate it to this repo myself, but I'd be happy to offer help to anyone who might want to take it on.

@muke101 Thank you very much for providing the PHAST implementation!

It does indeed have the potential to offer greater performance gains than the current StoreSet algorithm in XS-GEM5.

We will take a look at your implementation and try it out soon. Thank you.