Replace SYCL backend `reduce_by_segment` implementation with reduce-then-scan call#1915

Merged

mmichel11 merged 34 commits intomainfrom dev/mmichel11/rts_reduce_by_segment

Dec 19, 2024

+1,008-566

Commits on Dec 12, 2024

Initial commit of reduce_by_segment with the reduce-then-scan path
mmichel11
committed
Revert change to ranges and use zip_view over segments / values instead
mmichel11
committed
Implement correct return for reduce_by_segment
mmichel11
committed
Add support for flag predicates
mmichel11
committed
Revert "Add support for flag predicates"
mmichel11
committed
Re-implement support for flag predicates in a more performant manner
mmichel11
committed
Add fallback and remove old SYCL implementation
mmichel11
committed
Switch from using lambdas to functors
mmichel11
committed
Add device copyable specializations for red-by-seg functors and update testing
mmichel11
committed
Fix typo in error message in device_copyable.pass.cpp
mmichel11
committed
Introduce separate input generation for scan phase and update tests
mmichel11
committed
Improve code readability
mmichel11
committed
Add optional first key field to scan input and remove input range in write operations
mmichel11
committed
Update __write_op in reduce-then-scan
mmichel11
committed
Remove now unneeded ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro
mmichel11
committed
Alternate testing between usm shared and device to prevent excessive binary size
mmichel11
committed
Performance tuning within scan input functor
mmichel11
committed
Handle n=0, n=1 first in reduce_by_segment
mmichel11
committed
Code cleanup
mmichel11
committed
Improve comments and mark relevant variables as const
mmichel11
committed
Add condition to ensure value type is trivially copyable to call reduce-then-scan
mmichel11
committed
clang-format
mmichel11
committed
Introduce iterator based __pattern_reduce_by_segment
mmichel11
committed
Revert "Remove now unneeded ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro"
mmichel11
committed
Fix test bug where device allocation is always used for testing
mmichel11
committed
Separate each reduce_by_segment fallback path into their own functions
mmichel11
committed
clang-format
mmichel11
committed
Address comments in reduce-then-scan based implementation
mmichel11
committed
Improve explanations of reduce-by-segment approach
mmichel11
committed
Use binary_op[_non]_device_copyable where appropriate
mmichel11
committed
Address comments in fallback implementation
mmichel11
committed
Adjust comment to improve clarity

mmichel11
and
adamfidel
committed
Address review comment
mmichel11
committed
Use full namespace: oneapi::dpl::__ranges::__get_sycl_range to fix compilation issues
mmichel11
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Replace SYCL backend `reduce_by_segment` implementation with reduce-then-scan call#1915

Commits on Dec 12, 2024

Initial commit of reduce_by_segment with the reduce-then-scan path

Revert change to ranges and use zip_view over segments / values instead

Implement correct return for reduce_by_segment

Add support for flag predicates

Revert "Add support for flag predicates"

Re-implement support for flag predicates in a more performant manner

Add fallback and remove old SYCL implementation

Switch from using lambdas to functors

Add device copyable specializations for red-by-seg functors and update testing

Fix typo in error message in device_copyable.pass.cpp

Introduce separate input generation for scan phase and update tests

Improve code readability

Add optional first key field to scan input and remove input range in write operations

Update __write_op in reduce-then-scan

Remove now unneeded ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro

Alternate testing between usm shared and device to prevent excessive binary size

Performance tuning within scan input functor

Handle n=0, n=1 first in reduce_by_segment

Code cleanup

Improve comments and mark relevant variables as const

Add condition to ensure value type is trivially copyable to call reduce-then-scan

clang-format

Introduce iterator based __pattern_reduce_by_segment

Revert "Remove now unneeded ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro"

Fix test bug where device allocation is always used for testing

Separate each reduce_by_segment fallback path into their own functions

clang-format

Address comments in reduce-then-scan based implementation

Improve explanations of reduce-by-segment approach

Use binary_op[_non]_device_copyable where appropriate

Address comments in fallback implementation

Adjust comment to improve clarity

Address review comment

Use full namespace: oneapi::dpl::ranges::get_sycl_range to fix compilation issues