This page contains ideas for projects to work on during the gem5 coding sprint held in conjunction with HPCA 2017. For more information on the sprint, see: http://learning.gem5.org/tutorial/index.html.
Below are three categories of projects: small, medium, and large. I expect that only the "small" can be easily completed in an afternoon. For medium projects, you should be able to get a good start on them in an afternoon and finish up with a few extra hours of work after the sprint. Large project ideas are unlikely to be finished within a few days. These ideas are mainly for the community to get together and discuss at the sprint, not to start coding.
- 1 Small ideas
- 1.1 Fix TLB warmup for x86 [Jason]
- 1.2 Modify EventWrapper to understand C++11 lambdas [Jason]
- 1.3 Develop some ISA instruction tests [Many people]
- 1.4 Clean up serialization code [Andreas S]
- 1.5 CI smoke tests (faster than quick) [Andreas S]
- 1.6 Unify PS/2 handling [Andreas S]
- 1.7 Other ideas [Tony]
- 2 Medium sized projects
- 3 Long-term ideas to discuss
Fix TLB warmup for x86 [Jason]
See http://reviews.gem5.org/r/3474/. This patch was a first stab at fixing this problem, but it should be re-written and re-tested based on Steve's feedback.
The problem this patch tries to solve is that when switching from the atomic CPU to a timing CPU, the information in the TLB is lost. It is very important to not lose this TLB information when using the atomic CPU to warm up the system. By losing the TLB data, all experiments run with warmup will experience many more TLB misses than expected.
Note: This requires using FS mode to test.
Modify EventWrapper to understand C++11 lambdas [Jason]
The current EventWrapper can only wrap a very simple function that takes no parameters. However, I believe that C++11 supports passing pointers to partially bound function. See the documentation for std::bind: http://en.cppreference.com/w/cpp/utility/functional/bind.
With this feature, we should be able to delete most of the Event subclasses throughout the codebase, too. The idea is that most of the time the Event subclasses are only used to pass a single parameter from where the event is created to the event callback function.
This project should be quick for someone familiar with C++11 lambdas and early binding of parameters.
Develop some ISA instruction tests [Many people]
The goal of this project is to find out what is implemented correctly and possibly find some bugs in our ISAs. For instance, a recent GCC change exposed a bug in the x86 ISA.
The RISC-V insttest is an example of instruction tests that are already in the codebase. It should be pretty quick to put together a simple test which goes through some of the x86 and/or ARM instructions and explicitly tests each instruction like the RISC-V example above.
A longer-term goal would be to automate this process so we have some confidence that our implementation of the ISA is correct. Someone with experience validating hardware may have ideas on how to do this.
Clean up serialization code [Andreas S]
Increase the code reuse (particularly container helpers).
CI smoke tests (faster than quick) [Andreas S]
Create a separate test classification to quickly check to see if a change breaks anything major. Right now, contributors often just compile gem5 and say "close enough". This project would create a new class of tests (probably just move some tests around) so that developers can quickly (less than 5 minutes) test to see if they have made any breaking changes. This is also important for our upcoming CI changes.
Unify PS/2 handling [Andreas S]
PS2 is implemented by both the i8042 and PL050 models, but almost no code is shared. This should be an easy way to get your hands a little dirty with gem5's current code.
Other ideas [Tony]
- Adding checkpointing support to the GPU model
- Properly supporting atomics
- Add support for event-based scheduling in the GPU model, and FUPool-style functional units
Medium sized projects
- New test binaries based on the LLVM test suite [Andreas S]
- Mini-DSL for param overrides from the command line [Andreas S]
- Config cleanups. E.g., move some of config/common/ to a m5.config name space. [Andreas S]
- Proper support for pthreads in SE mode [Andreas S]
- Implement a fast mode in the HDLCD controller to support graphical workloads (e.g., Android) in KVM [Andreas S]
- Fixing the structure and design of the GPU coalescer [Tony]
Embed the generated system SVG in a web page [Andreas H]
It would be cool if you could interactively navigate through the system and see the simulation results.
This should be fairly easy for anyone skilled in client-side scripting. It may even be used to view incremental results while the simulation is running.
Long-term ideas to discuss
Re-write the test infrastructure [Jason]
Jason really doesn't like how the current test infrastructure uses SCons. I think it makes in incredibly hard to use. Plus, it makes it difficult to integrate into CI systems (e.g., Jenkins).
A couple of "requirements". This is not an exhaustive list.
- Incremental testing (don't test things for parts of the code that didn't change)
- Functional-only tests
- Easy to compare a subset of stats (e.g., all that matters is that the number of cache misses is the same)
Re-write the build system [Jason/Andreas S]
Andreas S. has taken some steps to replace SCons with cmake. Another interesting option is bazel: https://bazel.build/, a new build system from Google. This would be a rather huge undertaking, though.
A few requirements. Again, this list isn't exhaustive.
- Split gem5's components into separate dynamically-linked objects (reduce the size of the gem5 binary)
- Play nice with test and CI systems.
A binary-translation CPU [Andreas H]
A large-sized project for some crafty person out there. This would be useful for fast-forwarding, much like the KVMCpu, but more portable. It could, for example, be built on top of the Tiny Code Generator (TCG), as it is BSD licensed. Another possible place to look for inspiration is pydgin (https://github.com/cornell-brg/pydgin) which leverages Python (and PyPy/RPython) to build a jit for binary translation.
Quite a big task, but also a very big contribution to gem5.