rdaly525 / metamapper

Creating Mappers (synthesis tools) for CoreIR

MetaMapper

Creating Mappers (synthesis tools) for CoreIR

metamapper's Issues

Add ability to reconstruct the instruction from mapped json file

Once the application is mapped, there is no way to reconstruct the peak instruction from the mapped json file.

It would be great to add the ability to reconstruct the instruction since consumers other than the the one that generated the json can infer the application data. However, if this kind of behavior is intentional for some reasons, feel free to close the issue. I think I can do some hacks with the debug string.

WebAssembly to CoreiR

Hi @rdaly525,

My name is Guilherme and I am a PhD student at UFSCar (Brazil). I'm investigating ways to transform webassembly into CoreIR. Can your tool perform this operation?

My goal is to apply subgraph mining in WASM applications using GraMi. So, I think I'll need a CoreIR as input to GraMi.

Recreate match and replace pattern

Add Travis CI

Add Rewrite Rule construction from a peak simulation function (which uses multiple PEs)

It would be great to specify something like the following as a Rewrite rule (or library function)

def FMA(a : Data, b : Data, c : Data):
   x = PE(Inst.Mul,a,b)
   y = PE(Inst.Add,x,c)
   return y

Add a generic verify mapping (that each instance is a primitive) into the map_app

Not deleting temporary files causes buildup of temporary files

MetaMapper/metamapper/peak_util.py

Line 79 in b86e075

f = tempfile.NamedTemporaryFile(delete=False)

Is there a reason that delete=False? And is there a way of keeping track of these files and deleting them when they are not needed anymore?

Metamapper cannot map multiple applications currently.

This is due to the way that the instruction map is being generated.

Create a database for mapping result

The mapping takes multiple minutes per mapping on my local machine. Can you create a sqlite3 database after the mapping and publish it as a release file so that the downstream flow can use it? Ideally the PD tools can update the database with more accurate energy cost so that the mapper can choose how to map the op.

CI script to build cgralib

Retiming

We will need to ensure that operations that have non-zero latency are mapped properly. Other paths in the fabric must match the delay of operations that are a higher latency (e.g. ROM, floating point math).

Task List

• Figure out a plan to pass semantic information from Halide to CoreIR (What is Data, what is Valid, what is reset)
• There is an issue with getting into random states after configuration. Need to replace control registers to reset-able registers.
• Write a plan for how to specify pipelining/retiming

WASM tests are broken

Will depend on resolution of cdonovick/peak#208

Move lassen specific functions into lassen repo

Please add a LICENSE file

https://docs.github.com/en/free-pro-team@latest/github/building-a-strong-community/adding-a-license-to-a-repository

Enable loading lassen rewrite rules

Depends on StanfordAHA/lassen#185

Add ability for rewrite rules to require certain passes

Create an IO rewrite rule and IO primitive

Update rewrite rule synth to discover more (all) IR patterns

Given an IR rule, find all equivalent IR patterns
Will be Implemented using technique from Buchwald
This should solve any commutativity and associativity issues.