Skip to content

Current Project State

Project Scope

Workload

I'm aiming to support the risc-v vector spec. We need workloads that exercise the spec to ensure I implement it correctly, and I need workloads for common applications to check that I implement it with good performance. Currently I just have a few small kernels here.

Next Step: I have a FFT kernel that works in the python model. I can get some more quantitative measurements of the performance and see how it scales with the number of lanes. I need to better understand what the current bottleneck is.

Compiler

To get good performance I will need customizations to the compiler. I currently have a modification to LLVM to add a separate vector memory stack to support vector register spilling. It was a fairly small change and was done by Claude Code. I'll need to understand LLVM a lot better myself before I can start making changes I'm confident in.

Next Step: Current LLVM never uses unordered indexed vector operations. It would be cool to see if it could do that for cases where the lack of memory conflicts is easily inferred.

Architecture/Hardware Modelling

To be able to iterate a bit more quickly I have a python model of the hardware that captures the message passing approach. This covers a decent chunk of the vector spec now. Also, while updating the documentation I had lots of ideas for better ways to do things which I need to update the python model with.

Next Step: Continue to increase our coverage of the risc-v vector spec. Update to incorporate new approaches.

RTL

I'm slowly working on implementing the hardware design in Chisel. There is still a lot of work to do here.

Next Step: Get enough implemented that I can do basic vector loads and stores from the scalar memory and from the vector memory. This is mostly a case of implementing the Kamlet.

Backend

Currently I have a few of the modules from inside the jamlet running successfully through the backend flow using skywater130 at 100 MHz.

Next Step: Choose a module and get the building and using of hard macros working. Either the jamlet or kamlet would make most sense. It would also be good to get reasonable area number for the kamlet to make sure the kamlet overhead is not too expensive.