[SYCL][Docs] Add legacy SYCL 1.2.1 image aspect #9217
Conversation
Since SYCL 2020 no longer has the SYCL 1.2.1 images, neither the `device::info::image_support` query nor `aspect::image` query guarantees that a given device support the SYCL 1.2.1 images. To alleviate this, this commit adds an extension with a new aspect `ext_intel_legacy_image` intended for querying support for SYCL 1.2.1 images. To help users correctly transition to using this new aspect, a warning is issued when the user calls either `platform::has` or `device::has` with `aspect::image`, warning the user that SYCL 2020 images are not supported and suggesting they use the new aspect instead. In a future update, these calls will be changed to return false due to missing SYCL 2020 image support. Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>
|
Tag @intel/sycl-language-enabling-triage for awareness. |
| |=== | ||
|
|
||
|
|
||
| === More sections at your discretion |
There was a problem hiding this comment.
Need a better title for this section.
There was a problem hiding this comment.
Oops. It has been changed!
| #if !defined(SYCL_DISABLE_IMAGE_ASPECT_WARNING) && __has_attribute(diagnose_if) | ||
| #define __SYCL_WARN_IMAGE_ASPECT(aspect_param) \ | ||
| __attribute__((diagnose_if( \ | ||
| aspect_param == aspect::image, \ |
There was a problem hiding this comment.
What happens here if the value of aspect_param is not known at compile time?
There was a problem hiding this comment.
Sadly it would not produce the warning. For example a call like
sycl::aspect A = sycl::aspect::image;
Dev.has(A);would not produce the warning. However, the expectation is that using the aspect directly is the most common pattern and as such should catch the majority of cases.
There was a problem hiding this comment.
I think that's OK. I was worried that the compiler might diagnose an error if the expression aspect_param == aspect::image cannot be evaluated at compile time.
Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>
Signed-off-by: Larsen, Steffen <steffen.larsen@intel.com>
|
@intel/dpcpp-esimd-reviewers - In the original pass over the affected tests I missed the ESIMD-related ones. This has been amended in the newest merge with SYCL, but it should all just be requiring the new legacy image aspect instead of |
As mentioned in the CTS pull request KhronosGroup/SYCL-CTS#870, two device info descriptors, namely `info::device::image_max_array_size` and `info::device::opencl_c_version`, have been deprecated and no longer exist in the SYCL 2020 specification. But they still exist in our implementation, so I'm setting up deprecation warnings for them. Also a related test case has been revised to remove the usage of `info::device::opencl_c_version`. Also `info::device::usm_system_allocator` has been replaced and marked as deprecated 2+ years ago in PR #4007 and #9217 respectively. So I'm cleaning up a few relevant code pieces in our code as well in this PR. --------- Signed-off-by: Hu, Peisen <peisen.hu@intel.com>
Since SYCL 2020 no longer has the SYCL 1.2.1 images, neither the
device::info::image_supportquery noraspect::imagequery guarantees that a given device support the SYCL 1.2.1 images. To alleviate this, this commit adds an extension with a new aspectext_intel_legacy_imageintended for querying support for SYCL 1.2.1 images.To help users correctly transition to using this new aspect, a warning is issued when the user calls either
platform::hasordevice::haswithaspect::image, warning the user that SYCL 2020 images are not supported and suggesting they use the new aspect instead. In a future update, these calls will be changed to return false due to missing SYCL 2020 image support.