This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [GSoC] checking for the loop parallelism

On 31/07/2014 16:41, Roman Gareev wrote:
Tobias, could you please advise me how is it better to determine a
depth for loop (it is necessary for loop_is_parallel_p)? In
graphite-clast-to-gimple.c. it is done by using the number of
dimensions of a domain, which is attached to clast_for. As far as I
know, isl doesn't give such a opportunity. If I'm not mistaken â2 * (
level + 1) â is also not suitable for all cases.

Hi Roman,

you can get this information from the isl_ast_build that was used when generating a certain loop (you can access this isl_ast_build from the callbacks isl_ast_build_set_before_each_for and isl_ast_build_set_after_each_for). With isl_ast_build_get_schedule, you can get an incomplete schedule (less dimensions then the schedule that you gave to the isl ast generator). Specifically, it only contains the dimensions of the current loop and all surrounding ones. Consequently the last dimension in this incomplete schedule is the dimension you want to check for parallelism.

If you use this schedule in
loop_level_carries_dependences instead of the result from scop_get_transformed_schedule you should even be able to detect
parallelism for something like the following:

for (i
  A[i] += i

for (i)
  A[0] += i

Here the first loop is parallel, but the second not. The old approach would have said this dimension is not parallel at all. If you use the schedule provided, you only look at the statements involved in a certain loop, so the first loop would be detected as parallel.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]