[SYCL] Add implementation of kernel_bundle. Part 2 #3287
Conversation
The patch adds implementation of basic API and data structures. Error handling will added in a separate patch.
|
The clang-format check doesn't agree with clang-format on my machine. Suggest ignoring this fail. |
| detail::createSyclObjFromImpl<sycl::kernel_id>(KernleIDImpl); | ||
|
|
||
| // Insert new element keeping MKernelIDs sorted. | ||
| auto It = std::lower_bound(MKernelIDs.begin(), MKernelIDs.end(), KernelID, |
There was a problem hiding this comment.
Minor: I think pushing all the elements into the vector and then sorting it once should be more efficient than inserting each one while keeping the vector sorted.
There was a problem hiding this comment.
Will do measurements.
There was a problem hiding this comment.
The problem is that it might depends a lot on the real number of kernels.
Perhaps with SYCL generated by ML frameworks could have millions of kernels... ;-)
But for "normal" code and a few kernels?
Anyway, the std::sort path should be fast enough too for just a few kernels...
| @@ -0,0 +1,96 @@ | |||
| // RUN: %clangxx -fsycl -fsycl-device-code-split=per_kernel -fsycl-targets=%sycl_triple %s -o %t.out | |||
| // RUN: %t.out | |||
There was a problem hiding this comment.
According to post-commit testing this test doesn't work on the host device, so we either should move it to on-device directory or fix execution on the host device.
There was a problem hiding this comment.
The intention of the spec, of course, is that these APIs would work on all devices.
The patch adds implementation of basic API and data structures.
Error handling will added in a separate patch.