Classification into Application Domains
CPU-bound loops:
1. [CBD] Bandwidth Dependent, latency tolerant, FIR filter-like, loops without dependencies:
pipelined datapaths, digit-serial, parallelism by loop unrolling.
2. [CLD] Latency Dependent, IIR filter-like, dependent loops
- instruction level: conventional optimizations
- bit level: adapt arithmetic
- optimizations: constants, logic minimization, wired shift,…
- speciaI functions: e.g. 1/x,sin(x),max(*),…
Memory-bound loops:
3. [MBD] Memory Bandwidth Dependent:
minimize bits per data-value (requires “type size inference”)
4. [MLD] Memory Latency Dependent: e.g. pointers
implement data-structure + algorithm directly (CAM,Graph,BSAT)