sdgemm micro-kernel for the SG2042 RISC-V processor #868
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Hi, my name is Emanuele and I am a research fellow at the University of Bologna. In our lab, we are exploring the capabilities of emerging hardware platforms, including the SG2042 RISC-V processor. As part of this effort, also involving a student project, we selected BLIS.
As you may know, the SG2042 supports RVV 0.7.1, whereas BLIS currently includes kernels targeting RVV 1.0. Our goal was to make the RVV kernels compatible with the SG2042 using GCC 14. To achieve this, we performed a set of adaptations to target the xtheadvector machine architecture. Following an initial port, we conducted performance analysis and optimized the register grouping for this processor's register file, significantly reducing instruction count in the sdgemm micro-kernel.
Using HPL as a benchmark, we observed a 49% performance improvement over the standard BLIS RVV sdgemm micro-kernel.
Following your contribution guidelines, we integrated our work as a new configuration family: rvv_sg2042, which is fully compatible with GCC 14. The sdgemm micro-kernel has been optimized, while czgemm was adapted to compile and run correctly, though not performance-tuned.
We hope this contribution could be interesting for the BLIS community as well as the RISC-V community.