Question Just to make sure, if we have 8 program instances and if one of the instances takes much longer time than the other seven of them, do the other seven instances wait doing nothing until the slower instance is over?
This comment was marked helpful 0 times.
kayvonf
Good question. Conceptually yes. The semantics of the ISPC program are that when reaching the end of the ISPC function body, all instances wait until the entire gang has completed. Then control returns to the calling C code.
But your question seems to be more about the actual implementation, not the semantics. Now that you know that a gang of ISPC program instances is implemented as a sequence of SIMD instructions, and you know about SIMD execution, you are equipped to answer your question. Can someone take a shot?
One reason the instances might take different amounts of time to complete is that they execute different sequences of instructions (one instance might have more to do, and thus run longer). That is not the case in the above code as all instances will execution the execute same sequence. (Self-test: Do you see why this is true?)
However if the instances did execute different logic, and require different instruction sequences, as they may for the program written on this slide, then you can imagine an implementation might have to employ the techniques we discussed last lecture about mapping conditional control flow to SIMD hardware. In light of this implementation, can you now answer your question?
Question Just to make sure, if we have 8 program instances and if one of the instances takes much longer time than the other seven of them, do the other seven instances wait doing nothing until the slower instance is over?
This comment was marked helpful 0 times.
Good question. Conceptually yes. The semantics of the ISPC program are that when reaching the end of the ISPC function body, all instances wait until the entire gang has completed. Then control returns to the calling C code.
But your question seems to be more about the actual implementation, not the semantics. Now that you know that a gang of ISPC program instances is implemented as a sequence of SIMD instructions, and you know about SIMD execution, you are equipped to answer your question. Can someone take a shot?
One reason the instances might take different amounts of time to complete is that they execute different sequences of instructions (one instance might have more to do, and thus run longer). That is not the case in the above code as all instances will execution the execute same sequence. (Self-test: Do you see why this is true?)
However if the instances did execute different logic, and require different instruction sequences, as they may for the program written on this slide, then you can imagine an implementation might have to employ the techniques we discussed last lecture about mapping conditional control flow to SIMD hardware. In light of this implementation, can you now answer your question?
Also see http://ispc.github.com/ispc.html#control-flow
This comment was marked helpful 0 times.