Compiler Optimizations

• Reorder code to

  • Fuse loops
  • Combine aggregates
  • Combine broadcasts

loop1 X max

aggregate(maxX)

broadcast(maxX)

loop2 Y max

aggregate(maxY)

broadcast(maxY)

loop3 X min

aggregate(minX)

broadcast(minX)

loop4 Y min

aggregate(minY)

broadcast(minY)

loop X max, Y max, X min, Y min

aggregate(maxX, maxY, minX, minY)

broadcast(maxX, maxY, minX, minY)

• Code runs about 4 times faster

Notice: Such optimizations

result from the compiler’s

effectiveness, not directly

from the CTA model

Previous slide

Next slide

Back to first slide

View graphic version