[SYCL] Add Experimental Range Rounding

[hdelan]

This commit adds new clang command line option -fsycl-exp-range-rounding. Experimental range rounding maps all 1, 2 or 3 dim range kernels to range rounded kernels which can be rounded in all dims.

SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=x:y:z can be used for exp range rounding, where the y param will be a factor of the rounded range in all dims

[hdelan]

@intel/dpcpp-clang-driver-reviewers I am not sure about compiler flags. Instead of introducing the new -fsycl-force-range-rounding, should I instead make a new flag -fsycl-range-rounding=[force/disable] which would deprecate the existing -fsycl-disable-range-rounding?

[mdtoguchi]

@intel/dpcpp-clang-driver-reviewers I am not sure about compiler flags. Instead of introducing the new -fsycl-force-range-rounding, should I instead make a new flag -fsycl-range-rounding=[force/disable] which would deprecate the existing -fsycl-disable-range-rounding?

I like the idea of introducing a new -fsycl-range-rounding=<arg> option. It alleviates any potential confusion when multiple range-rounding options are used on the command line.

[aelovikov-intel]

+1 to -fsycl-range-rounding=[disable|on|force]. I'd also split the experimental 1D rounding into a separate PR.

[hdelan]

@mdtoguchi @aelovikov-intel changes for the new -fsycl-range-rounding flag are here. I will keep this PR for the experimental range rounding

[rafbiels]

I tested this PR with https://gist.github.com/rafbiels/2b584584cfd6412e6b255adab4c264d6 on NVIDIA sm_86 GPU. Here's the outcomes:

   Release     Time [ms]     CUDA gridXYZ     CUDA blockXYZ
-----------------------------------------------------------
  2023.2.1            63     {2, 1888, 1}       {394, 1, 1}
  2024.0.2           223      {788, 3, 1}       {1, 768, 1}
     test1            66     {197, 16, 1}       {4, 118, 1}
     test2            61     {25, 118, 1}       {32, 16, 1}

test1 = hdelan:1d-range-rounding with default flags
test2 = hdelan:1d-range-rounding with -fsycl-exp-range-rounding

The improvement between 2024.0.2 and test1 is thanks to #12663 but we can see that this PR (test2) improves the local range assignment further and the performance surpasses the earlier naive implementation from 2023.2.1.

[hdelan]

Ping @againull @intel/llvm-reviewers-runtime @intel/dpcpp-cfe-reviewers @intel/dpcpp-clang-driver-reviewers @intel/unified-runtime-reviewers

[aelovikov-intel]

I tested this PR with https://gist.github.com/rafbiels/2b584584cfd6412e6b255adab4c264d6 on NVIDIA sm_86 GPU. Here's the outcomes:

Can we add it into test-e2e/PerformanceTests ?

[elizabethandrews]

FE changes LGTM

[hdelan]

The perf tests stuff is in a very early stage, we're basically finding it out as we go. Currently it's just the logs captured for the tests there even in case of a PASS, the rest is left to humans to interpret.

OK sure thanks

[aarongreig]

approving to satisfy the requirement but I don't think UR reviewers are responsible for any of the code changed

[hdelan]

Actually I just realized there are no FE tests. Please add a test for the predefines.

Test added, thanks.

[mdtoguchi]

OK for Driver

[smanna12]

FE changes LGTM

[hdelan]

@intel/llvm-gatekeepers this can be merged now