[SYCL] Use correct units when specifying image offset and extent #19936
Conversation
AlexeySachkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
AlexeySachkov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
AlexeySachkov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
AlexeySachkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
AlexeySachkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
AlexeySachkov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
| // ur_rect_offset_t and ur_rect_region_t describe their first component as | ||
| // bytes, whilst ze_image_region_t uses pixels. | ||
| // | ||
| // However, the getImageRegionHelper above is used for both bindless and regular |
There was a problem hiding this comment.
Note for UR reviewers: this inconsistency in the spec was reported as URT-969
|
Was this not covered by any test? Could we add one now? |
Yes and no. We do have tests for copies, but a lot of them UPD: to be precise, out of 6 bindless image copy tests, 4 have The change itself is |
…correct-units-in-bindless-image-copies
AlexeySachkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
AlexeySachkov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
AlexeySachkov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
| impl->MDestOffset = {DestOffset[0], DestOffset[1], DestOffset[2]}; | ||
| // Copy args computed here are directly passed to UR. Various offsets and | ||
| // extents end up passed as ur_rect_offset_t and ur_rect_region_t. Both those | ||
| // structs expect theirfirst component to be in bytes, not in pixels |
There was a problem hiding this comment.
| // structs expect theirfirst component to be in bytes, not in pixels | |
| // structs expect their first component to be in bytes, not in pixels |
|
@intel/unified-runtime-reviewers-level-zero, could you please take a look? I'm mostly looking for feedback about the new helper function that I'm introducing. I can't modify existing one, because there are other APIs which use it, but they expect the "old" The way of how the helper will be used isn't final, because |
urBindlessImagesImageCopyExpdocuments its argument that describe copy region as usingbytes-pixel-pixelformat, while we were mistakenly passingpixel-pixel-pixelformat in there.This PR adjusts that. It is an
NFCchange for CUDA & HIP adapters, because for them it only changes the place where pixel to byte conversion happens, but the change is tricker for L0 adapter. In L0, there are APIs which expectpixel-pixel-pixelformat and there are APIs which acceptbytes-pixel-pixelformat, so some extra care is needed there.