Even though superscalar processors silently take advantage of ILP and provide speed ups without users needing to know what's happening underneath the hook, programmers can willingly increase the ILP of their programs to increase the effectiveness of these processors. Increasing ILP is often times easier to do compared to program-wise parallelism. Small modifications to existing programs could increase performance significantly. One example technique is loop unrolling, which is explained in the 213 book (chapter 5). This paper gives another detailed explanation between ILP and loop unrolling for those interested.
This comment was marked helpful 1 times.
jpaulson
@Xiao: I think usually these types of optimizations are best left to the compiler. Your source code doesn't have much resemblance to the assembly that is output by sophisticated compilers, and they actively look for these opportunities.
This comment was marked helpful 0 times.
martin
I believe that ILP is part of the micro-architecture. So exploit of parallelism using techniques such as branch predictor, out-of-order execution underneath doesn't effect programmer level interface. And there are many techniques that programmer can use including loop unrolling as @xiao says.
Even though superscalar processors silently take advantage of ILP and provide speed ups without users needing to know what's happening underneath the hook, programmers can willingly increase the ILP of their programs to increase the effectiveness of these processors. Increasing ILP is often times easier to do compared to program-wise parallelism. Small modifications to existing programs could increase performance significantly. One example technique is loop unrolling, which is explained in the 213 book (chapter 5). This paper gives another detailed explanation between ILP and loop unrolling for those interested.
This comment was marked helpful 1 times.
@Xiao: I think usually these types of optimizations are best left to the compiler. Your source code doesn't have much resemblance to the assembly that is output by sophisticated compilers, and they actively look for these opportunities.
This comment was marked helpful 0 times.
I believe that ILP is part of the micro-architecture. So exploit of parallelism using techniques such as branch predictor, out-of-order execution underneath doesn't effect programmer level interface. And there are many techniques that programmer can use including loop unrolling as @xiao says.
This comment was marked helpful 0 times.