- PlanNode recursive IR -> optimization IR produce equivalent or more efficent IR to avoid having to decide between different plans.
- New Operations NOP, Interval
- ProjFuse: fusing table scan proj -> bottom left TS -> PJ fusing together on right side. enables parse selected columns for each row
- Pushdown Proj: partly implement 4th lecture -> cooperates well with last pass.
- Pred Pass: merge Pred on same column -> equivalent expr to replace both ops with singe op or check contradict -> this case contradict not leq 3 and geq 25 -> compiler replaces with NOP and effectivly discardes querry
2.
- Perf adaptive SIMD ops against optimizing compiler with scalar ops -> both equiv performance -> due to small dataset size, overhead from unpacking records from vec dominates overall computation
- Tested on gpu-a100 Icelake, our SIMD ops more outliers than AMD Genoa -> diff of microarchitecture
J: Welcome to our presentation of the results of course work,
we will start by quickly outlining the architecture of our query engine