[SYCL] Honor dependencies of empty command groups

[aelovikov-intel]

No description provided.

[aelovikov-intel]

Not very familiar with the scheduler, so this can be totally wrong. And if this is the right direction, then there might be some work left for @intel/sycl-graphs-reviewers to support the same when part of SYCL graph.

[EwanC]

Not very familiar with the scheduler, so this can be totally wrong. And if this is the right direction, then there might be some work left for @intel/sycl-graphs-reviewers to support the same when part of SYCL graph.

Thanks for the ping, it's a good question. I think we already have a test for this which hasn't regressed based on the CI of this PR.
Our add_empty_node() E2E helper function does

  Graph.begin_recording(Queue);
  auto ev = Queue.submit(
      [&](sycl::handler &CGH) { CGH.depends_on(std::vector<event>{Deps...}); });
  Graph.end_recording(Queue);

Which should be a code path stressed by this input https://github.com/intel/llvm/blob/sycl/sycl/test-e2e/Graph/Inputs/empty_node.cpp used as part of E2E test https://github.com/intel/llvm/blob/sycl/sycl/test-e2e/Graph/RecordReplay/empty_node.cpp

[sergey-semenov]

LGTM, just a tiny nitpick.

[sarnex]

@aelovikov-intel Are these postcommit fails related?

https://github.com/intel/llvm/actions/runs/12036585964/job/33558551326

# |   what():  Enqueue process failed.

[aelovikov-intel]

Hm, I wouldn't expect. I'm going to wait till the job finishes and restart the failing task plus see how the next commit's post-commit CI behaves.

[aelovikov-intel]

Interesting,

#include <sycl/sycl.hpp>

int main() {
  sycl::queue q;
  sycl::buffer<int, 1> b{1};

  q.submit([&](sycl::handler &cgh) {
    sycl::accessor a{b, cgh, sycl::read_write_host_task};
  });
  return 0;
}

passes, but

#include <sycl/sycl.hpp>

using namespace sycl::access;

int main() {
  sycl::queue q;

  sycl::unsampled_image<1> UImg(sycl::image_format::r32b32g32a32_uint, 3);

  q.submit([&](sycl::handler &cgh) {
    auto A1 =
        UImg.get_access<sycl::uint4, mode::read, sycl::image_target::host_task>(
            cgh);
  });
  return 0;
}

fails:

$ clang++ -fsycl a.cpp -O0 -gdwarf-4
$ ./a.out
terminate called after throwing an instance of 'sycl::_V1::exception'
  what():  Enqueue process failed.
Aborted (core dumped)

[sarnex]

I don't understand why it didn't fail in precommit either. It looks like the failure case is L0 + gen12 + linux, which we definitely have in precommit. Weird.