About
Early reports of compiling cactuBSSN_r with rv64gcv indicate vectorization may be making this code worse. The report indicates we're executing ~114b additional instructions. This needs to be understood and addressedCactuBSSN was showing a 1-2% degradation in instruction counts with vector enabled. That has been fixed as of b7b387e1200f on the GCC trunk and we see roughly a 1% dynamic instruction count improvement.
Beyond fixing the regression, vector is not expected to provide much benefit for CactuBSSN. Testing on the k1 chip shows that 1.2% improvement in dynamic instruction counts using vector, but a 1% performance regression. As noted in the cam4 work item, we believe this is due to weaknesses in the k1 vector unit design.
Stakeholders/Partners
RISE:
Ventana: Jeff Law
External:
Dependencies
Status
| Page Properties | ||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
Updates
- Add data from run on the k1 (BPI-F3 board)
- Regression has been fixed on the GCC trunk
- Competitive data shows no significant improvement with vector
- The 1% reduction in instruction counts we see with vector enabled is not likely to move the needle in any significant way performane-wise
- Considering this done/closed.
- Project reported as a priority for 1H2024
...