[SYCL][CUDA] Fix context scope in kernel launch #4606
Conversation
The `guessLocalWorkSize` function uses the CUDA API so it needs an active context, and there was no active `ScopedContext` when it was called which may cause issue. This fixes intel#2777
There was a problem hiding this comment.
LGTM.
Beware that #4401 also implements this by putting the ScopedContext inside guessLocalWorkSize. I do not have a preference to whether or not it should be at API top-level or inside guessLocalWorkSize, but having it both places is redundant. Ping @JackAKirk for awareness.
|
Looks fine to me. You might want to wrap the ScopedContext into a try/catch block in case it throws PI_INVALID_CONTEXT, that can then be returned as pi_result in cuda_piEnqueueKernelLaunch; since this appears to be the standard practice in the file. |
Good catch, I think we should try to keep the
Good point, I'll extend the |
npmiller
dismissed stale reviews from JackAKirk and steffenlarsen
via
55f2f4f
Signed-off-by: JackAKirk <jack.kirk@codeplay.com>
* upstream/sycl: (2344 commits) [ESIMD] Rename slm_load4/slm_store4 to slm_load_rgba/slm_store_rgba (intel#4158) [SYCL] Avoid re-computing group_range in nd_item::get_group_range() (intel#4621) [clang-offload-extract] Ignore zero padding in .tgting section (intel#4622) [Driver][SYCL] Fix -fsycl-help output when redirected (intel#4619) [Driver][SYCL][FPGA] Do not unbundle aoco as an archive for hardware (intel#4477) [Driver][SYCL] Fix offload-bundler and offload-deps triples (intel#4616) [SYCL] Fix bit_cast for half type (intel#4603) [SYCL] Fix a typo in accessor::get_range method (intel#4556) [SYCL] Detach allocas from their dependencies regardless of linked alloca presence (intel#4573) [SYCL][L0] Make sure that we only query/sync host-visible events from the host. (intel#4613) Fix tests with wrong alias metadata [Driver][SYCL] Fixup arguments to llc call for PIC and code-model (intel#4614) [SYCL][L0] Add ownership control for LeveL-Zero kernel_bundle interop. (intel#4576) [SYCL][Driver] Expose llvm-foreach --jobs functionality through a driver option (intel#4543) [SYCL] Prevent stream buffer leak on constructor exception (intel#4594) [ESIMD] Replace mask_type_t with simd_mask to represent Gen predicates. (intel#4230) Fix for a bunch of fixed point integer SPIR-V instructions (intel#1213) Amend SingleElementVectorINTEL decoration use cases according to spec update (intel#1192) Enforce UserSemantic decoration if no FPGA decorations found [SYCL][CUDA] Fix context scope in kernel launch (intel#4606) ...
The
guessLocalWorkSizefunction uses the CUDA API so it needs anactive context, and there was no active
ScopedContextwhen it wascalled which may cause issue.
This fixes #2777