PPT Slide
2. Precision / Range Optimization
How many bits do we need to represent a variable?
Why? (a) move floating point code to fix point
(b) reduce required memory bandwidth
(c) minimize required FPGA area
(d) lower latency / higher throughput
(e) lower power consumption
- Integer Code + - *: Compiler Analysis
Bitwise[MIT], Constantinides[Imperial], Razdan[Harvard]
- Floating Point, including Division, is a harder problem.
=> Combine Interval Arithmetic & Automatic Differentiation