[SYCL] Refactor SYCL kernel object handling in hierarchical parallelism #6212
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -65,6 +65,19 @@ | |
| // (1) - materialization of a PFWI object | ||
| // (2) - "fixup" of the private variable address. | ||
| // | ||
| // TODO: add support for the case when there are other functions between | ||
bader marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| // parallel_for_work_group and parallel_for_work_item in the call stack. | ||
| // For example: | ||
| // | ||
| // void foo(sycl::group<1> group, ...) { | ||
| // group.parallel_for_work_item(range<1>(), [&](h_item<1> i) { ... }); | ||
| // } | ||
| // ... | ||
| // cgh.parallel_for_work_group<class kernel>( | ||
| // range<1>(...), range<1>(...), [=](group<1> g) { | ||
| // foo(g, ...); | ||
| // }); | ||
| // | ||
| // TODO The approach employed by this pass generates lots of barriers and data | ||
| // copying between private and local memory, which might not be efficient. There | ||
| // are optimization opportunities listed below. Also other approaches can be | ||
|
|
@@ -385,16 +398,8 @@ static void copyBetweenPrivateAndShadow(Value *L, GlobalVariable *Shadow, | |
| LocAlign = MaybeAlign(AI->getAlignment()); | ||
| } else { | ||
| auto Arg = cast<Argument>(L); | ||
| T = Arg->getParamByValType(); | ||
| LocAlign = MaybeAlign(Arg->getParamAlignment()); | ||
| } | ||
|
|
||
| assert(T && "Unexpected type"); | ||
|
|
@@ -698,16 +703,7 @@ static void fixupPrivateMemoryPFWILambdaCaptures(CallInst *PFWICall) { | |
| // Go through "byval" parameters which are passed as AS(0) pointers | ||
| // and: (1) create local shadows for them (2) and initialize them from the | ||
| // leader's copy and (3) materialize the value in the local variable before use | ||
| static void shareByValParams(Function &F, const Triple &TT) { | ||
| // Skip alloca instructions and split. Alloca instructions must be in the | ||
| // beginning of the function otherwise they are considered as dynamic which | ||
| // can cause the problems with inlining. | ||
|
|
@@ -726,29 +722,20 @@ static void sharePFWGPrivateObjects(Function &F, const Triple &TT) { | |
| Instruction &At = LeaderBB->back(); | ||
|
|
||
| for (auto &Arg : F.args()) { | ||
| if (!Arg.hasByValAttr()) | ||
|
Contributor
There was a problem hiding this comment. Nit: we skip "this" because it is allocated in the proper AS by the FE, correct? Comment would be helpful for the reader.
Contributor
Author
There was a problem hiding this comment. Right. I just reverted the changes from #1455 and tried to re-implement it by fixing address space in clang instead. |
||
| continue; | ||
|
|
||
| assert(Arg.getType()->getPointerAddressSpace() == | ||
| asUInt(spirv::AddrSpace::Private)); | ||
|
|
||
| // Create the shared copy - "shadow" - for current arg | ||
| Type *T = Arg.getParamByValType(); | ||
| GlobalVariable *Shadow = | ||
| spirv::createWGLocalVariable(*F.getParent(), T, "ArgShadow"); | ||
|
|
||
| LLVMContext &Ctx = At.getContext(); | ||
| IRBuilder<> Builder(Ctx); | ||
| Builder.SetInsertPoint(&LeaderBB->front()); | ||
|
|
||
| copyBetweenPrivateAndShadow(&Arg, Shadow, Builder, | ||
| true /*private->shadow*/); | ||
|
|
@@ -766,6 +753,7 @@ PreservedAnalyses SYCLLowerWGScopePass::run(Function &F, | |
| FunctionAnalysisManager &FAM) { | ||
| if (!F.getMetadata(WG_SCOPE_MD)) | ||
| return PreservedAnalyses::all(); | ||
| LLVM_DEBUG(llvm::dbgs() << "Function name: " << F.getName() << "\n"); | ||
| const auto &TT = llvm::Triple(F.getParent()->getTargetTriple()); | ||
| // Ranges of "side effect" instructions | ||
| SmallVector<InstrRange, 16> Ranges; | ||
|
|
@@ -866,9 +854,8 @@ PreservedAnalyses SYCLLowerWGScopePass::run(Function &F, | |
| for (auto *PFWICall : PFWICalls) | ||
| fixupPrivateMemoryPFWILambdaCaptures(PFWICall); | ||
|
|
||
| // Finally, create shadows for and replace usages of byval pointer params. | ||
| shareByValParams(F, TT); | ||
|
|
||
| #ifndef NDEBUG | ||
| if (HaveChanges && Debug > 0) | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.