[SYCL] Get rid of device kernel info duplication #20319
Conversation
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Error
sergey-semenov
force-pushed
the
pmduplication
branch
from
47f5f86 to
d70616d
Compare
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
force-pushed
the
pmduplication
branch
from
d70616d to
b85a646
Compare
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Error
With the introduction of DeviceKernelInfo, assert usage and implicit local argument information is now duplicated in program manager. This patch removes the duplicate maps and makes it so that device kernel info map is filled out during image registration, with the compile time information added when it's available (during the first submission of the kernel).
sergey-semenov
force-pushed
the
pmduplication
branch
from
b85a646 to
3f0ce59
Compare
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
| return MOwnsDeviceKernelInfo | ||
| ? MDeviceKernelInfo | ||
| : ProgramManager::getInstance().getDeviceKernelInfo( |
There was a problem hiding this comment.
Can we just have a reference/pointer as the member (potentially pointing to the owning smart pointer member) instead of doing those lookups?
There was a problem hiding this comment.
Yes, we should. I was planning that as a follow-up change since the lookup is already there right now.
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
is known and unrelated. |
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
|
@intel/sycl-graphs-reviewers (@mmichel11), @intel/llvm-reviewers-runtime (@cperkinsintel) Could you please have a look? |
mmichel11
left a comment
mmichel11
left a comment
There was a problem hiding this comment.
Changes to CommandGraph tests LGTM
|
@cperkinsintel gentle ping |
| CompileTimeKernelInfoTy DefaultCompileTimeInfo{std::string_view(name)}; | ||
| m_DeviceKernelInfoMap.try_emplace(std::string_view(name), | ||
| DefaultCompileTimeInfo); |
There was a problem hiding this comment.
IIUC, name is a local variable and we are using its string_view as key to m_DeviceKernelInfoMap. Could this lead to dangling pointers in the map when local variable goes out of scope?
There was a problem hiding this comment.
name is a const char* taken from the device image here, so there won't be any problems unless the entry outlives the module that contains that image.
Now, there is a more general issue here if we, say, add images A and B coming from different libraries and containing the same kernel name, and then drop just the image we took the string_view from. But that's not introduced by this PR and I'm planning to address that soon.
| bool kernel_impl::checkOwnsDeviceKernelInfo() { | ||
| // If the image originates from something other than standard offline | ||
| // compilation, this kernel needs to own its info structure. | ||
| // We could also have a mixed origin image, in which case the device kernel |
There was a problem hiding this comment.
Sorry, could you please elaborate on how can we have a mixed origin image?
There was a problem hiding this comment.
For example, if 2 images with different origins are explicitly linked, or an image has a dependency on a devicelib image.
|
@intel/llvm-gatekeepers please consider merging |
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
sergey-semenov
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
4 participants
With the introduction of DeviceKernelInfo, implicit local argument information is now duplicated in program manager. This patch removes the duplicate map and makes it so that device kernel info map is filled out during image registration, with the compile time information added when it's available (during the first submission of the kernel).
Additinally, this patch ensures that device kernel info instances are added to the program manager only during image registration, and gets rid of entries from other sources that polluted the map.