Didn’t write them out. 😔
What Do I Do?
I aim to make hardware design and compiler design as joyful as software design can be. Beauty and joy are valid motivations in science and engineering; if they’re not, let my career founder.
What Am I Writing?
I’m writing a blog post of my unfinished LATTE draft. I’m also writing up a Glenside blog post to go with the web demo.
Chris Lattner’s ASPLOS 2021 Keynote
Love the way he describes how software people are pulled in. Happens, if we’re lucky, near the start of the hardware design process, but often in the middle.
He’s also describing how an accelerator gets more and more complex as the design evolves. Specifically the software layer expanding. Very relatable. Feels like it’s leading to some good punchline, but we’ll see if our ideas diverge.
I have vague questions here about how we handle the system-level concerns. Also how we handle data movement.
“Most of this [referring to complex architectures] is not related to the target workload.” Yes! Most of the time and effort is spent on the glue.
Use open standards! Yes, agree, that’s what we did in the software/OS world. We’ve standardized busses already. Great.
Control processor is often something that people will write by hand for fun, at a lot of cost (security, power management, lots of bugs, etc). RISC-V, unsurprisingly ;)
Fleshing out some of my thoughts: so much of “heterogeneous computing” is still actually “CPU-dependent” or “CPU-centric” computing. Everything centers around the CPU and the “main” memory hierarchy. The most trivial questions, like how and when do I move data in and out of my accelerator, become absolutely central.
“Hardware design is ripe for opportunities to…reevaluate.”
My question: “In my limited experience (mostly in the TVM world), a lot of compiler concerns are actually system-level concerns: moving data on and off device, dispatching commands to those devices, etc. These concerns are often handled with some proprietary runtime, that, as a compiler person, I often like to ignore as not fitting into the clean abstractions compilers provide. Do you see MLIR (or some other future solutions) as eventually providing a unification for those system-level concerns?” His answer: the runtime is definitely the elephant in the room. Runtimes contain secret sauce, so they can’t be shared and people don’t want to unify them. We need some kind of dual to MLIR in the runtime space. He named one from Google, I think?
This feeds more deeply into the sense I have that, put roughly, the software world has C/Linux, and the hardware world has nothing. Processes, ELF/C calling interface, and syscalls are the low-level, generally-accepted interfaces in the software world. Do we have that in the hardware world?
Review of Progress 2021
We do a review of progress each year in our PhD program. Seems like a good thing to put in my research journal. Putting it online makes it higher stakes :)
What are your goals for the next year?
- I’ve submitted a paper for publication in the past few weeks (meeting my goal from 2020), but regardless of whether that gets accepted or rejected, I’d like to have another paper accepted within the year.
- Make actionable connections at my Google internship, hopefully leading to another internship or research collaboration in the hardware design/architecture/compilers/PL space; e.g. with the CIRCT folks.
What are your long-term career goals? My answer from last year hasn’t changed: I’d still like to go into industry after grad school, and maybe return to academia later in my life (likely not in a high-powered R1-school PI role, though). However, I think much less about the distant future these days, given the pandemic.
Name one thing we (advisor or department) could do better to help you reach these goals. Same answer as last year…connections to industry are definitely the most important thing for me.
When do you plan to complete your next milestone? At my current pace (especially considering I’m going on internship), winter/spring 2022 seems like a reasonable target for generals, which is right in keeping with the department’s schedule. I don’t know what makes me ready for generals, though. I have a rough idea of the research direction for the rest of my PhD (compiler techniques to aid in hw/sw codesign) but haven’t worked towards this. I might do some work along these lines in the coming weeks before my internship, which would be awesome, but I suspect that the 590 I’m taking plus wrapping up RTML phase 1 will take most of my time.
In reviewing your goals for this past year, how did your accomplishments measure up to your planned goals?
My goals from last year:
A year from now, I see myself:
- Having submitted or having a plan to submit applications for industry fellowships
- Having “completed” (i.e. written a paper for) a project on hardware-software codesign for ML accelerators. (The project might be ongoing, but by then I expect at least the first “chapter” to be complete and its findings summarized in a paper!)
- Having submitted a paper to ASPLOS 2021 or ISCA 2021 on the above project (and it hopefully having been accepted!)
- Mentoring at least one undergraduate student on some project – whether it be related to my current work or to my quals project work
- On the horizon (probably for 2021): Generals, plus I’d love to talk about teaching a class!
No plans for industry fellowships. I could go for QIF next year; I think I’d be well-positioned for it. I’m not really sure on other industry fellowships. I’ll probably need to sprint for them in September of this year.
Wrote and submitted a paper for hw/sw codesign (or, that was the project’s original intent!)
Did not submit to ISCA or ASPLOS this year; both times were my decision. I wasn’t happy with the story we were telling. The story we ended up telling with our recent submission felt much more genuine. I feel I made the right call.
Mentored Andrew Liu, and started working with Vishal Canumalla! There’s no shortage of great undergrads.
See previous answer on generals. A year ago, I was hopeful to be ahead of schedule (and also wanting that pay bump!), but I feel the time (and money) pressure less and less. I like being here :)
What happens if you run verification in reverse?
Is this what synthesis is? Starting from a specification and searching?
Will there someday be analogues of the Halide algorithm–schedule split for hardware? Where we split hardware functionality from specific concerns i.e. timing, handshake protocols, etc.?
Luis Vega recommended this to me.
“Compiler optimizers solve a search problem, but without doing any search.”
Markov Chain Monte Carlo: what is it? Make a random change; if the cost got better, always accept; if the cost increases, with some probability accept anyway. Is this a general description of MCMC?
The collection of 6 operations that he describes can be used to implement all compiler optimizations! Constant propagation, dead code elimination, etc. Also they’re not minimal; in MCMC, sometimes you want to take big steps, sometimes you want to take small steps.
Two-phase process: first, search for correct programs, then search for optimizations on top of those.
But what is “cost”, when we’re looking for correctness? First pass: number of incorrect bits on the outputs of a test suite. This doesn’t work exactly right; instead, reward getting the right answer in the wrong place (i.e. register).
The “stair-step” pattern is typical. It goes for a long time and suddenly makes a big improvement. Interested to see how he explains that, because on its face, it’s really frustrating!
Vector increment example is interesting.
It’s making me wonder:
is there always some obvious
to achieving some algorithmic goal?
Actually he just said,
if your task
has this structure,
then this approach can work.
Someone asks: why not start from the correct program? A great question. His answer is basically: there’s no reason to assume that the correct program is good. If we’re looking for a better program, and we start from the correct program, we will never want to leave that correct program, because the correctness is so enticing! It’s better to start from nothing and find correctness “in the wild”; starting from a local minimum of correctness is a good way to never leave that local minimum! Going off that: are there better starting places than a single return statement?
They do a slow test using SMT to check equivalence.
Similarly they have fast and slow tests for performance. Fast tests just add latencies of instructions; slow tests benchmark code in a sandbox.
How do we determine two programs are equal? There’s no provenance—i.e. the sequence of rewrites getting you from the original to the final—so we’re stuck trying to answer one of the harder verification problems of two arbitrary programs being equal. Well if they’re loop free, then we can do it with SMT.
But what about loopy code? We can break the proof into subproofs, where each subproof relies only on loop-free code. But how do we generate those subproofs? We match the program at “cutpoints”, which are guessed-and-checked by running the code and seeing how many times each line of code is run, and correlating them. This is interesting!
Instead of reading the paper, find a video of the presentation! I can’t believe it’s taken me this long to realize this. Basically anything is preferable to reading the paper. Find a video, read a blog post, read the abstract, and only after all of that isn’t enough, then read the paper.
Here’s a question I’d like to explore in my PhD: How can we have the hardware designer take a more active role in compiler design?
In my limited understanding of the relationship between hardware d
ASPLOS 2021 Better Hardware Through Compilers Session
Watching Clement’s Cuttlesim talk.
Synthesis and simulation toolchains should be separate.
Cuttlesim generates sequential C++
It also makes efforts
to make generated code readable,
which is a big plus.
There’s also an anecdote about using
to make hardware performance counters.
Also cool to see that a paper like this
is welcomed at ASPLOS!
Rachit’s Calyx talk!
We have new high-level abstractions
in HDL world.
Everyone has to build
the entire compiler stack.
Like LLVM, but LLVM doesn’t capture
structure and time.
Like FIRRTL, but
FIRRTL doesn’t capture control flow.
Has a control flow language
of the hardware graph
can be referred to by name.
Next talk is on virtualization of FPGAs. This was a best paper, but I honestly tuned out! I feel bad that this subject isn’t interesting to me.
I want to build hardware tools that spark joy when using them. That is a good description of my motivation.
Trying to get MLIR building:
cmake -G Ninja ../llvm \ -DLLVM_ENABLE_PROJECTS=mlir \ -DLLVM_BUILD_EXAMPLES=ON \ -DLLVM_TARGETS_TO_BUILD="X86;NVPTX;AMDGPU" \ -DCMAKE_BUILD_TYPE=Release \ -DLLVM_ENABLE_ASSERTIONS=ON \ -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DLLVM_ENABLE_LLD=ON
(As per these instructions.)
Running into this
cmake configuration issue
CMake Error at cmake/modules/HandleLLVMOptions.cmake:291 (message): Host compiler does not support '-fuse-ld=lld'
lld is optional,
I’ll try it without that flag.
Got it building, put it in a Docker: https://github.com/gussmith23/2021-04-20-mlir-experiment.
I’m starting to move towards FPGA stuff. I have to refresh my knowledge on what the internals of FPGAs look like. I have a PYNQ board on my desk, so I’m looking through this data sheet: https://www.xilinx.com/support/documentation/data_sheets/ds190-Zynq-7000-Overview.pdf. (Mine has an XC7Z020 on it; see Table 1 on page 3.)
Reflection on Goals
Didn’t write any goals, but had a very productive week in the end.