Remove `make_iter_mode` and refactor access mode resolution by danhoeflinger · Pull Request #2551 · uxlfoundation/oneDPL

danhoeflinger · 2025-12-29T17:26:23Z

Summary

Move access mode resolution logic from scattered make_iter_mode calls to the leaf-level __get_sycl_range when processing sycl_iterator. This centralizes the mode resolution logic and ensures it properly recurses through nested iterator types.

Changes

Removed make_iter_mode pattern:

Removed iter_mode struct and make_iter_mode function from parallel_backend_sycl.h
Removed all make_iter_mode wrapper calls from algorithm implementations in algorithm_impl_hetero.h and async_impl_hetero.h

Refactored iter_mode_resolver:

Moved to utils_ranges_sycl.h as the authoritative location
Primary template is now undefined (SFINAE-friendly) instead of using static_assert
Added __is_iter_mode_resolvable_v trait for testable compile-time compatibility checking
Added third template parameter noInit to track whether algorithm allows skipping host→device copy

Fixed mode compatibility logic:

Removed incorrect specializations that claimed read or write iterators could satisfy read_write algorithm requirements
Removed incorrect specializations allowing discard_* modes with algorithms that dont have no_init
Valid combinations: exact match, read_write→read, read_write→write, discard_read_write -> * (only when no_init=true), discard_write -> write (only when no_init=true).

Connected no_init property to accessors:

all_view now includes a template argument for _NoInit which connects to actual sycl accessor
Now properly uses backend buffer accessor system with no_init property when appropriate

NOTE: These changes could result in new build error for users where it was not before if it was relying upon improperly matched behavior. It seems like this could've resulted in silent problems or even segfaults in kernels where algorithms don't get the access modes they need.

Added tests:

test_is_iter_mode_resolvable_v() covering valid and invalid mode combinations

Targeted to land after #2519 and #2549 (final episode in the __get_sycl_range refactor trilogy)
Resolves Issue #2550

Future work:
There may be some opportunity to utilize no_init for our temporary buffers, though I'm not sure if that will result in any real savings, as the system should know when there is nothing to transfer to the device.

Copilot

Pull request overview

This PR centralizes access mode resolution logic for SYCL iterators by moving it from scattered make_iter_mode calls to the leaf-level __get_sycl_range implementation. The refactoring removes the iter_mode struct and make_iter_mode function, replacing them with compile-time mode resolution directly in __process_input_iter for sycl_iterator types.

Key Changes:

Removed make_iter_mode pattern and moved access mode resolution into __get_sycl_range's sycl_iterator specialization
Fixed incorrect mode compatibility logic (e.g., read/write iterators can no longer satisfy read_write algorithm requirements)
Added comprehensive test coverage for mode resolution compatibility checking

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
`parallel_backend_sycl.h`	Removed deprecated `iter_mode_resolver`, `iter_mode` struct, and `make_iter_mode` function (~130 lines)
`utils_ranges_sycl.h`	Added refactored `iter_mode_resolver` with corrected compatibility logic and new `sycl_iterator` specialization in `__process_input_iter`
`algorithm_impl_hetero.h`	Removed `make_iter_mode` wrapper calls from fill, generate, partition_copy, inplace_merge, and partial_sort patterns
`async_impl_hetero.h`	Removed `make_iter_mode` wrapper calls from async fill pattern
`get_sycl_range.pass.cpp`	Added `test_is_iter_mode_resolvable_v` test function covering valid/invalid mode combinations

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

MikeDvorskiy · 2026-01-12T16:35:40Z

At first glance, I am in favor of the changes. I didn’t like iter_mode/iter_mode_resolver 3–4 years ago, since I had noticed needless structures. I’ve been wanting to cut these structures out of the oneDPL code for a long time.

Thank you!

SergeyKopienko · 2026-01-13T10:02:16Z

                //try searching for the first element which not equal to *__b
                if (__b != __first1)
-                    __b += __internal::__pstl_upper_bound(__b, _DifferenceType1{0}, __last1 - __b, __b, __comp, __proj1, __proj1);
+                    __b += __internal::__pstl_upper_bound(__b, _DifferenceType1{0}, __last1 - __b, __b, __comp, __proj1,


Let's save the previous format.

removed these formatting only changes. thanks.

SergeyKopienko · 2026-01-13T10:02:20Z

                //try searching for the first element which not equal to *__e
                if (__e != __last1)
-                    __e += __internal::__pstl_upper_bound(__e, _DifferenceType1{0}, __last1 - __e, __e, __comp, __proj1, __proj1);
+                    __e += __internal::__pstl_upper_bound(__e, _DifferenceType1{0}, __last1 - __e, __e, __comp, __proj1,


Let's save the previous format.

removed these formatting only changes. thanks.

fix access modes for uninitialized apis remove unnecessary pattern Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

Co-authored-by: Alexey Kukanov <alexey.kukanov@intel.com>

…emove_make_iter_mode

MikeDvorskiy

Left a comment/question.

…ake_iter_mode

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com1>

akukanov · 2026-03-13T11:12:16Z

NOTE: These changes could result in new build error for users where it was not before if it was relying upon improperly matched behavior. It seems like this could've resulted in silent problems or even segfaults in kernels where algorithms don't get the access modes they need.

In other words, it is a potential breaking change for some users. We should not do this without analysis and discussion. And if we decide to do it in the future, we'll need to bump the major library version and perhaps even update the specification.

danhoeflinger · 2026-03-13T12:19:10Z

NOTE: These changes could result in new build error for users where it was not before if it was relying upon improperly matched behavior. It seems like this could've resulted in silent problems or even segfaults in kernels where algorithms don't get the access modes they need.

In other words, it is a potential breaking change for some users. We should not do this without analysis and discussion. And if we decide to do it in the future, we'll need to bump the major library version and perhaps even update the specification.

Yes, I suppose it is a breaking change, for users who had their access mode "hints" completely ignored and overridden. Both are valid implementations of the spec I believe, but you are right that it could break users code that currently works.
I'll revert those narrowing changes, and restore the previous access mode resolver which only rejects write with a read hint and vice versa.

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

danhoeflinger · 2026-03-13T18:53:46Z

In the review process, we have come to the conclusion that this should be considered a bit more robustly before merging. User provided hints are really only even considered for a small subset of algorithms currently. This PR would make them be resolved with algorithmic needs everywhere.

Also, it seems like the benefit of being stricter on conforming to user supplied hints is not worth the breaking changes that come from it due to poor coverage of using make_iter_mode prior to this PR.

It seems the best plan of action may be to consider hints for for_each and related APIs, but ignore them in all other cases.
for_each patterns could get some performance gain from a user providing more information about the usage pattern, allowing the implementation to skip copies to or from the device. All other APIs have known algorithmic needs, so hints don't make much sense, and should likely be ignored (which is allowed by the spec).

I have part of this proposal on a branch implemented, but it requires some more significant changes than can happen before code freeze, so I'm taking it out of the milestone. We lose out on connecting no_init to sycl buffer accessors, but we can include this in the next release.

danhoeflinger · 2026-03-30T12:27:16Z

closing in favor of #2617.

danhoeflinger requested review from MikeDvorskiy, SergeyKopienko and Copilot December 29, 2025 17:26

Copilot AI reviewed Dec 29, 2025

View reviewed changes

danhoeflinger force-pushed the dev/dhoeflin/remove_make_iter_mode branch from e4040d0 to 5bf81d6 Compare December 31, 2025 21:50

danhoeflinger requested a review from Copilot January 2, 2026 12:56

Copilot AI reviewed Jan 2, 2026

View reviewed changes

danhoeflinger requested review from akukanov, dmitriy-sobolev, mmichel11 and timmiesmith January 2, 2026 15:37

danhoeflinger linked an issue Jan 2, 2026 that may be closed by this pull request

iter_mode / make_iter_mode issues and potential cleanup #2550

Open

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from 36ae926 to 03c83d1 Compare January 2, 2026 22:33

danhoeflinger force-pushed the dev/dhoeflin/remove_make_iter_mode branch from 044c431 to 728d20c Compare January 2, 2026 22:35

danhoeflinger mentioned this pull request Jan 5, 2026

Refactor __get_sycl_range to align with sycl #2519

Merged

danhoeflinger added this to the 2022.12.0 milestone Jan 5, 2026

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from 03c83d1 to 08eb860 Compare January 5, 2026 18:12

danhoeflinger force-pushed the dev/dhoeflin/remove_make_iter_mode branch from 808f0de to be5fe30 Compare January 5, 2026 18:13

SergeyKopienko reviewed Jan 13, 2026

View reviewed changes

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from 08eb860 to f02042d Compare January 13, 2026 13:50

danhoeflinger force-pushed the dev/dhoeflin/remove_make_iter_mode branch from be5fe30 to c0563c6 Compare January 13, 2026 14:01

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from f02042d to e029ce4 Compare February 10, 2026 16:40

danhoeflinger force-pushed the dev/dhoeflin/remove_make_iter_mode branch from 267d03c to 1b39713 Compare February 10, 2026 17:07

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from e029ce4 to 46ff098 Compare February 20, 2026 16:41

danhoeflinger mentioned this pull request Feb 25, 2026

Access mode changes for transform_if #2556

Merged

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from 3060a9d to 90ff4fe Compare February 25, 2026 16:06

danhoeflinger added 2 commits February 25, 2026 15:23

refactor away redundant fill functor;

ec2d9db

fix access modes for uninitialized apis remove unnecessary pattern Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

fixing typo uninitialized_value_construct

d2e7ad7

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

danhoeflinger and others added 8 commits March 2, 2026 11:05

formatting fixes

76cdfd3

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

formatting

0f229d5

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

adding memory_impl_hetero.h

ed00490

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

formatting, naming

36f57e0

fix header

5364964

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

remove stale sycl_traits

47cffba

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

using fully qualified name

89aec8e

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

fix header copyright

0a56ec7

Co-authored-by: Alexey Kukanov <alexey.kukanov@intel.com>

danhoeflinger force-pushed the dev/dhoeflin/unitialized_refactors branch from 0b61bcf to 0a56ec7 Compare March 2, 2026 16:05

Merge branch 'dev/dhoeflin/unitialized_refactors' into dev/dhoeflin/r…

394ea4a

…emove_make_iter_mode

MikeDvorskiy reviewed Mar 6, 2026

View reviewed changes

Comment thread include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h

danhoeflinger mentioned this pull request Mar 6, 2026

Investigate converting ranges API usages of views::all_write to write with no_init #2599

Open

Base automatically changed from dev/dhoeflin/unitialized_refactors to main March 11, 2026 17:10

danhoeflinger added 2 commits March 11, 2026 14:26

Merge remote-tracking branch 'origin/main' into dev/dhoeflin/remove_m…

1e5cbe7

…ake_iter_mode

::std -> std

c1efdff

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

akukanov reviewed Mar 12, 2026

View reviewed changes

Comment thread include/oneapi/dpl/pstl/memory_impl.h Outdated

Comment thread include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h Outdated

Comment thread include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h Outdated

address feedback

b20b931

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

akukanov reviewed Mar 12, 2026

View reviewed changes

Comment thread include/oneapi/dpl/pstl/hetero/dpcpp/utils_ranges_sycl.h Outdated

removing useless comments

a11e729

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com1>

danhoeflinger added 2 commits March 13, 2026 11:40

update to restore previous funcitonality from iter_mode_resolver

2a2396a

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

formatting

bafe7f0

Signed-off-by: Dan Hoeflinger <dan.hoeflinger@intel.com>

danhoeflinger modified the milestones: 2022.12.0, 2022.13.0 Mar 13, 2026

danhoeflinger closed this Mar 30, 2026

danhoeflinger deleted the dev/dhoeflin/remove_make_iter_mode branch March 30, 2026 12:27

akukanov removed this from the 2022.13.0 milestone Apr 20, 2026

Conversation

danhoeflinger commented Dec 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

MikeDvorskiy commented Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SergeyKopienko Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

danhoeflinger Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

SergeyKopienko Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

danhoeflinger Jan 13, 2026

Choose a reason for hiding this comment

Uh oh!

MikeDvorskiy left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

akukanov commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

danhoeflinger commented Mar 13, 2026

Uh oh!

danhoeflinger commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

danhoeflinger commented Mar 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

danhoeflinger commented Dec 29, 2025 •

edited

Loading

MikeDvorskiy commented Jan 12, 2026 •

edited

Loading

akukanov commented Mar 13, 2026 •

edited

Loading

danhoeflinger commented Mar 13, 2026 •

edited

Loading