Goals

(didn’t fill out any for this week)

What Do I Do?

I’m trying to modify an environment variable in the middle of a Python script. I set os.environ["LD_LIBRARY_PATH"] in the script, but it doesn’t seem to be “seen” by the processes launched later on. I’ve confirmed that this is the problem, as when I run the same Python script with LD_LIBRARY_PATH=<value> prepended, where <value> is the value that should be getting set in the script, it works. Setting it even more explicitly in the script to exactly what I want is also not working. I think this is my own misunderstanding of environment variables and os.environ. It may also be some TVM weirdness in how TVM is handling environment variables.

As far as I can tell, up til this point, the changes to LD_LIBRARY_PATH are being seen. When I print out LD_LIBRARY_PATH just before that line, I see the path I expect. So it must be that this version of the variable is not being used?

This post is making me realize that this may just be a fundamental limitation of Python.

I thought I might use pytest-env, but even that doesn’t seem to be working. This is frustrating!

It seems like the problem is that the dynamic linker is initialized long before any of these environment variables are being set. I’m going to try building a static library instead.

That worked! Sad that I couldn’t figure out a different solution here, but I think this is fine in the end.

Lying in bed last night, I had some realizations.

A common one, which has come to me before: I control completely what I do. If there’s an idea, nobody’s going to tell me to work on it, or give me the time to work on it. I need to do that myself.

Another is that I think I should stay attached, weekly, to what my larger goals and ideas are.

Another, related, is that if there’s an idea I have, there is no better way of evaluating the idea than directly working on it. There is no faster and surer way to know whether something will work than by just building the thing. Moreover, research inspiration seems (for me) to come only from getting my hands dirty. It doesn’t come from some lofty place. It comes from writing code and not knowing what you’re doing, and slowly coming to know what you’re doing, and how it could be done better.

And again, related: I am the biggest gatekeeper to myself. Access to hardware and software are potential future problems, of course, but at the moment, I’m the one stopping myself from working on ideas that I have. For example, if I want to work on FPGAs, as I have for a while, there is literally nothing stopping me but myself. I just grabbed an FPGA from my building and set it up last week. I can do whatever I want with it.

I’m putting some time in to the experiment I’ve been working off-and-on on for a while now. It’s a new version of Glenside that has two primary new features:

  1. It’s layout independent. Tensor types are no longer tuples; instead, they are maps from dimensions IDs to positive integers.
  2. It has a more complex type system. Specifically, the root input nodes are equal if they have the same shape, not if they have the same value. The way I built the first version of Glenside made it so that I couldn’t actually represent similar networks overlapping in the egraph, because the inputs would all be assumed to be unique. This necessitates adding an apply node, that actually applies an expression on a given set of tensors to produce a result.

Working on these two ideas simultaneously may muddle things. I might need to work on just one at a time. Idea 1 is more likely to produce something interesting immediately, like a layout search. Idea 2 is more interesting from the egraphs perspective, and more interesting to me.

What do I want to work on for my molecular computing class? I took Luis’s DNA class last year at this time. I worked on an egraphs-based DNA strand displacement (DSD) project. The goal was to mess around with egraphs and egg. It also tied in with this larger goal I have of exploring some general idea of compiling workloads to computational primitives. My takeaway was that, if anything like this could ever exist, we’ll have to somehow overcome the fact that it will constantly be very domain-specific. The idea in general is that we have workloads—things we want to compute—and we also have compute primitives, provided, in short, by physics. Transistor-based logic, commonly, but also molecular computing, quantum computing, etc. I have this idea in the back of my head that we could somehow catalog these primitives and then, given a workload, somehow search over this library of primitives to figure out how we should compute the thing.

Week Recap

This week ended up being eaten up by a last minute deadline push. But we got the thing in! Very proud, no matter what happens. Learned a lot, and felt validated by the process.