Bindless image desc match ur

[cperkinsintel]

The bindless image image_descriptor does not have pitch/slice. We need that to properly create a ur_image_desc_t . Without it we'll never be able to support linear images that have aligned strides. There are also some older SYCL 1.2.1 defensive checks that will also block.

Added a unit test that demonstrates that these values correctly arrive at the UR.

Here is the ur_image_desc_t from ur_api.h in case anyone is wondering:

typedef struct ur_image_desc_t {
  /// [in] type of this structure, must be ::UR_STRUCTURE_TYPE_IMAGE_DESC
  ur_structure_type_t stype;
  /// [in][optional] pointer to extension-specific structure
  const void *pNext;
  /// [in][nocheck] memory object type
  ur_mem_type_t type;
  /// [in] image width
  size_t width;
  /// [in] image height
  size_t height;
  /// [in] image depth
  size_t depth;
  /// [in] image array size
  size_t arraySize;
  /// [in] image row pitch
  size_t rowPitch;
  /// [in] image slice pitch
  size_t slicePitch;
  /// [in] number of MIP levels, must be `0`
  uint32_t numMipLevel;
  /// [in] number of samples, must be `0`
  uint32_t numSamples;

} ur_image_desc_t;

[Copilot]

Pull request overview

This PR extends the SYCL bindless image image_descriptor so that layout-related fields (row pitch / slice pitch, plus num_samples) can be carried through to Unified Runtime (UR) ur_image_desc_t, enabling correct interop with externally-provided image memory that uses aligned/strided layouts.

Changes:

• Extend sycl::ext::oneapi::experimental::image_descriptor with row_pitch, slice_pitch, and num_samples, and thread these through UR descriptor construction.
• Update bindless image UR-struct population to use descriptor-provided pitches (when an explicit pitch argument is not provided).
• Add unit tests and documentation updates covering the new descriptor fields.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
sycl/unittests/Extensions/BindlessImages/ImageDescriptors.cpp	Adds UR-mock-based unit tests asserting descriptor layout fields arrive at UR.
sycl/unittests/Extensions/BindlessImages/CMakeLists.txt	Registers the new unit test source file.
sycl/source/handler.cpp	Updates UR image descriptor filling for bindless image copy command setup.
sycl/source/detail/image_impl.hpp	Adjusts checkImageDesc declaration (removes unused UserPtr parameter).
sycl/source/detail/image_impl.cpp	Adjusts checkImageDesc definition/call site and removes SYCL 1.2.1-era defensive checks.
sycl/source/detail/bindless_images.cpp	Propagates descriptor-provided row/slice pitch and num_samples into UR image descriptors.
sycl/include/sycl/ext/oneapi/bindless_images_descriptor.hpp	Adds layout fields to the public descriptor type and updates constructors accordingly.
sycl/doc/extensions/experimental/sycl_ext_oneapi_bindless_images.asciidoc	Documents the new descriptor fields/constructors.

[cperkinsintel]

ping to reviewers. The L0 team needs this for their work on linear/DRM

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

[iclsrc]

Round 3 follow-up review. The latest commits correctly address previous feedback: num_samples, row_pitch, and slice_pitch are now fully propagated through both the populate_ur_structs (bindless path) and fill_image_desc/fill_copy_args (handler path) routes. The "explicit pitch arg wins when non-zero" precedence rule is consistently implemented and the new unit tests cover the key propagation scenarios. The removal of redundant validations in checkImageDesc is safe because getImageDesc already guarantees those invariants before calling the checker. No Critical or Important issues found.

[dyniols]

@cperkinsintel Let me know if you add more changes. For now except few comments it looks good.

[dyniols]

Wouldn't be more reasonable to make 1 as default value since single sample == 1?

[cperkinsintel]

I think we should keep it at 0. That matches how its siblings ( row_pitch, slice_pitch etc behave). To my mind that clearly denotes "we aren't multi-sampling" and we can derive it if needed. Note that if we did change it, we'd also have to change constructor default values, including for constructors that someone might write in the future. Much safer to follow the initialize to zero norm.

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.