Skip to content

SYCL: skip grid-stride loop when not needed for MDRange#9250

Draft
masterleinad wants to merge 2 commits into
kokkos:developfrom
masterleinad:sycl_grid_stride_loop
Draft

SYCL: skip grid-stride loop when not needed for MDRange#9250
masterleinad wants to merge 2 commits into
kokkos:developfrom
masterleinad:sycl_grid_stride_loop

Conversation

@masterleinad

@masterleinad masterleinad commented Jun 5, 2026

Copy link
Copy Markdown
Contributor

Analogous to #9142 for SYCL. Copilot wrote the code since I wanted to try it and it seemed simple enough. I did some cleanup.

before:

1: OverlapMDRangePolicy/N:200/M:10000/R:10                                                                                                                                        0.058 s         0.058 s            12       0.0106696     5.34192m               0.0101212                  5.08093m    5.08321m          5.041u
1: MDRangeStencil_2D_MDRange_LayoutRight/size:512/tile_size:0/iterations:1/manual_time                                                                                            0.076 ms        0.078 ms            1              1          8         32
1: MDRangeStencil_3D_MDRange_LayoutLeft/size:128/tile_size:0/iterations:1/manual_time                                                                                             0.093 ms        0.094 ms            1              1         32          2          4
1: MDRangeStencil_4D_MDRange_LayoutRight/size:32/tile_size:0/iterations:1/manual_time                                                                                             0.073 ms        0.074 ms            1              1          2          2          2         16
      Start  4: Kokkos_PerformanceTest_MDRangePolicy_Stream
4: Test command: /home/darndt/kokkos/build/core/perf_test/Kokkos_PerformanceTest_MDRangePolicy_Stream "--benchmark_counters_tabular=true" "--benchmark_out=Kokkos_PerformanceTest_MDRangePolicy_Stream_2026-06-05_T19-10-24.json"
4: Running /home/darndt/kokkos/build/core/perf_test/Kokkos_PerformanceTest_MDRangePolicy_Stream
4: MDRangePolicy_Set<1>/22/manual_time        0.508 ms        0.508 ms         1529 1.78503k/s    907.039
4: MDRangePolicy_Set<3>/22/manual_time        0.613 ms        0.613 ms         1221 1.48012k/s    907.039
4: MDRangePolicy_Set<6>/22/manual_time         2.84 ms         2.84 ms          247  319.439/s    907.039
4: MDRangePolicy_Triad<1>/22/manual_time       1.38 ms         1.38 ms          510 1.96657k/s   2.72112k
4: MDRangePolicy_Triad<3>/22/manual_time       1.62 ms         1.62 ms          429 1.68467k/s   2.72112k
4: MDRangePolicy_Triad<6>/22/manual_time       3.81 ms         3.81 ms          185  714.687/s   2.72112k
 4/12 Test  #4: Kokkos_PerformanceTest_MDRangePolicy_Stream ...   Passed    6.58 sec

after:

1: OverlapMDRangePolicy/N:200/M:10000/R:10                                                                                                                                        0.058 s         0.058 s            12       0.0104729      5.3311m               0.0101351                  5.07832m    5.07214m          5.248u
1: MDRangeStencil_2D_MDRange_LayoutRight/size:512/tile_size:0/iterations:1/manual_time                                                                                            0.063 ms        0.064 ms            1              1          8         32
1: MDRangeStencil_3D_MDRange_LayoutLeft/size:128/tile_size:0/iterations:1/manual_time                                                                                             0.078 ms        0.078 ms            1              1         32          2          4
1: MDRangeStencil_4D_MDRange_LayoutRight/size:32/tile_size:0/iterations:1/manual_time                                                                                             0.079 ms        0.080 ms            1              1          2          2          2         16
      Start  4: Kokkos_PerformanceTest_MDRangePolicy_Stream
4: Test command: /home/darndt/kokkos/build/core/perf_test/Kokkos_PerformanceTest_MDRangePolicy_Stream "--benchmark_counters_tabular=true" "--benchmark_out=Kokkos_PerformanceTest_MDRangePolicy_Stream_2026-06-05_T19-10-24.json"
4: Running /home/darndt/kokkos/build/core/perf_test/Kokkos_PerformanceTest_MDRangePolicy_Stream
4: MDRangePolicy_Set<1>/22/manual_time        0.521 ms        0.521 ms         1492 1.74044k/s    907.039
4: MDRangePolicy_Set<3>/22/manual_time        0.640 ms        0.640 ms         1120 1.41771k/s    907.039
4: MDRangePolicy_Set<6>/22/manual_time         2.07 ms         2.07 ms          344  439.084/s    907.039
4: MDRangePolicy_Triad<1>/22/manual_time       1.38 ms         1.38 ms          511 1.96756k/s   2.72112k
4: MDRangePolicy_Triad<3>/22/manual_time       1.60 ms         1.60 ms          439 1.69945k/s   2.72112k
4: MDRangePolicy_Triad<6>/22/manual_time       3.09 ms         3.09 ms          227  880.916/s   2.72112k
 4/12 Test  #4: Kokkos_PerformanceTest_MDRangePolicy_Stream ...   Passed    6.25 sec

So we see quite some improvement for MDRangeStencil_3D_MDRange_LayoutLeft, MDRangePolicy_Triad<6>andMDRangePolicy_Set<6>/22/manual_time` while the other times stay more or less the same.

masterleinad and others added 2 commits June 5, 2026 11:27
@masterleinad
SYCL: skip grid-stride loop when not needed for MDRange (align with PR …
95ef08e 
…kokkos#9142)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Daniel Arndt <arndtd@ornl.gov>
@masterleinad
Clean-up
fdf6370 
Signed-off-by: Daniel Arndt <arndtd@ornl.gov>
yasahi-hpc
yasahi-hpc approved these changes Jun 8, 2026
View reviewed changes

@yasahi-hpc yasahi-hpc left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@Adrien-Tab

Adrien-Tab commented Jun 8, 2026

Copy link
Copy Markdown
Member

Just a note that for the other backends (CUDA and HIP), the bandwidth gap between ranks 3 and 6 is less significant

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Adrien-Tab Adrien-Tab Adrien-Tab approved these changes

@yasahi-hpc yasahi-hpc yasahi-hpc approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@masterleinad @Adrien-Tab @yasahi-hpc