[mlir][OpenMP] Added translation for omp.teams to LLVM IR
#68042
Conversation
This file contains 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 patch adds translation from `omp.teams` operation to LLVM IR using OpenMPIRBuilder. The clauses are not handled in this patch.
shraiysh
requested review from
skatrak,
jsjodin,
kiranchandramohan,
raghavendhra,
DominikAdamski,
dpalermo and
kiranktp
|
@llvm/pr-subscribers-mlir-openmp @llvm/pr-subscribers-flang-openmp ChangesThis patch adds translation from Full diff: https://github.com/llvm/llvm-project/pull/68042.diff 2 Files Affected:
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 8f7f1963b3e5a4f..b9643be40e13c01 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -661,6 +661,24 @@ convertOmpSingle(omp::SingleOp &singleOp, llvm::IRBuilderBase &builder,
return bodyGenStatus;
}
+// Convert an OpenMP Teams construct to LLVM IR using OpenMPIRBuilder
+static LogicalResult convertOmpTeams(omp::TeamsOp op, llvm::IRBuilderBase &builder, LLVM::ModuleTranslation &moduleTranslation) {
+ using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
+ LogicalResult bodyGenStatus = success();
+ if(op.getNumTeamsLower() || op.getNumTeamsUpper() || op.getIfExpr() || op.getThreadLimit() || !op.getAllocatorsVars().empty() || op.getReductions()) {
+ return op.emitError("unhandled clauses for translation to LLVM IR");
+ }
+ auto bodyCB = [&](InsertPointTy allocaIP, InsertPointTy codegenIP){
+ LLVM::ModuleTranslation::SaveStack<OpenMPAllocaStackFrame> frame(moduleTranslation, allocaIP);
+ builder.restoreIP(codegenIP);
+ convertOmpOpRegions(op.getRegion(), "omp.teams.region", builder, moduleTranslation, bodyGenStatus);
+ };
+
+ llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
+ builder.restoreIP(moduleTranslation.getOpenMPBuilder()->createTeams(ompLoc, bodyCB));
+ return bodyGenStatus;
+}
+
/// Converts an OpenMP task construct into LLVM IR using OpenMPIRBuilder.
static LogicalResult
convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
@@ -2406,6 +2424,9 @@ LogicalResult OpenMPDialectLLVMIRTranslationInterface::convertOperation(
.Case([&](omp::SingleOp op) {
return convertOmpSingle(op, builder, moduleTranslation);
})
+ .Case([&](omp::TeamsOp op) {
+ return convertOmpTeams(op, builder, moduleTranslation);
+ })
.Case([&](omp::TaskOp op) {
return convertOmpTaskOp(op, builder, moduleTranslation);
})
diff --git a/mlir/test/Target/LLVMIR/openmp-teams.mlir b/mlir/test/Target/LLVMIR/openmp-teams.mlir
new file mode 100644
index 000000000000000..c9005fca94a7c20
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-teams.mlir
@@ -0,0 +1,136 @@
+// RUN: mlir-translate -mlir-to-llvmir -split-input-file %s | FileCheck %s
+
+llvm.func @foo()
+
+// CHECK-LABEL: @omp_teams_simple
+// CHECK: call void {{.*}} @__kmpc_fork_teams(ptr @{{.+}}, i32 0, ptr [[wrapperfn:.+]])
+// CHECK: ret void
+llvm.func @omp_teams_simple() {
+ omp.teams {
+ llvm.call @foo() : () -> ()
+ omp.terminator
+ }
+ llvm.return
+}
+
+// CHECK: define internal void @[[outlinedfn:.+]]()
+// CHECK: call void @foo()
+// CHECK: ret void
+// CHECK: define void [[wrapperfn]](ptr %[[global_tid:.+]], ptr %[[bound_tid:.+]])
+// CHECK: call void @[[outlinedfn]]
+// CHECK: ret void
+
+// -----
+
+llvm.func @foo(i32) -> ()
+
+// CHECK-LABEL: @omp_teams_shared_simple
+// CHECK-SAME: (i32 [[arg0:%.+]])
+// CHECK: [[structArg:%.+]] = alloca { i32 }
+// CHECK: br
+// CHECK: [[gep:%.+]] = getelementptr { i32 }, ptr [[structArg]], i32 0, i32 0
+// CHECK: store i32 [[arg0]], ptr [[gep]]
+// CHECK: call void {{.+}} @__kmpc_fork_teams(ptr @{{.+}}, i32 1, ptr [[wrapperfn:.+]], ptr [[structArg]])
+// CHECK: ret void
+llvm.func @omp_teams_shared_simple(%arg0: i32) {
+ omp.teams {
+ llvm.call @foo(%arg0) : (i32) -> ()
+ omp.terminator
+ }
+ llvm.return
+}
+
+// CHECK: define internal void [[outlinedfn:@.+]](ptr [[structArg:%.+]])
+// CHECK: [[gep:%.+]] = getelementptr { i32 }, ptr [[structArg]], i32 0, i32 0
+// CHECK: [[loadgep:%.+]] = load i32, ptr [[gep]]
+// CHECK: call void @foo(i32 [[loadgep]])
+// CHECK: ret void
+// CHECK: define void [[wrapperfn]](ptr [[global_tid:.+]], ptr [[bound_tid:.+]], ptr [[structArg:.+]])
+// CHECK: call void [[outlinedfn]](ptr [[structArg]])
+// CHECK: ret void
+
+// -----
+
+llvm.func @my_alloca_fn() -> !llvm.ptr<i32>
+llvm.func @foo(i32, f32, !llvm.ptr<i32>, f128, !llvm.ptr<i32>, i32) -> ()
+llvm.func @bar()
+
+// CHECK-LABEL: @omp_teams_branching_shared
+// CHECK-SAME: (i1 [[condition:%.+]], i32 [[arg0:%.+]], float [[arg1:%.+]], ptr [[arg2:%.+]], fp128 [[arg3:%.+]])
+
+// Checking that the allocation for struct argument happens in the alloca block.
+// CHECK: [[structArg:%.+]] = alloca { i1, i32, float, ptr, fp128, ptr, i32 }
+// CHECK: [[allocated:%.+]] = call ptr @my_alloca_fn()
+// CHECK: [[loaded:%.+]] = load i32, ptr [[allocated]]
+// CHECK: br label
+
+// Checking that the shared values are stored properly in the struct arg.
+// CHECK: [[conditionPtr:%.+]] = getelementptr {{.+}}, ptr [[structArg]]
+// CHECK: store i1 [[condition]], ptr [[conditionPtr]]
+// CHECK: [[arg0ptr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 1
+// CHECK: store i32 [[arg0]], ptr [[arg0ptr]]
+// CHECK: [[arg1ptr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 2
+// CHECK: store float [[arg1]], ptr [[arg1ptr]]
+// CHECK: [[arg2ptr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 3
+// CHECK: store ptr [[arg2]], ptr [[arg2ptr]]
+// CHECK: [[arg3ptr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 4
+// CHECK: store fp128 [[arg3]], ptr [[arg3ptr]]
+// CHECK: [[allocatedPtr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 5
+// CHECK: store ptr [[allocated]], ptr [[allocatedPtr]]
+// CHECK: [[loadedPtr:%.+]] = getelementptr {{.+}}, ptr [[structArg]], i32 0, i32 6
+// CHECK: store i32 [[loaded]], ptr [[loadedPtr]]
+
+// Runtime call.
+// CHECK: call void {{.+}} @__kmpc_fork_teams(ptr @{{.+}}, i32 1, ptr [[wrapperfn:@.+]], ptr [[structArg]])
+// CHECK: br label
+// CHECK: call void @bar()
+// CHECK: ret void
+llvm.func @omp_teams_branching_shared(%condition: i1, %arg0: i32, %arg1: f32, %arg2: !llvm.ptr<i32>, %arg3: f128) {
+ %allocated = llvm.call @my_alloca_fn(): () -> !llvm.ptr<i32>
+ %loaded = llvm.load %allocated : !llvm.ptr<i32>
+ llvm.br ^codegenBlock
+^codegenBlock:
+ omp.teams {
+ llvm.cond_br %condition, ^true_block, ^false_block
+ ^true_block:
+ llvm.call @foo(%arg0, %arg1, %arg2, %arg3, %allocated, %loaded) : (i32, f32, !llvm.ptr<i32>, f128, !llvm.ptr<i32>, i32) -> ()
+ llvm.br ^exit
+ ^false_block:
+ llvm.br ^exit
+ ^exit:
+ omp.terminator
+ }
+ llvm.call @bar() : () -> ()
+ llvm.return
+}
+
+// Check the outlined function.
+// CHECK: define internal void [[outlinedfn:@.+]](ptr [[data:%.+]])
+// CHECK: [[conditionPtr:%.+]] = getelementptr {{.+}}, ptr [[data]]
+// CHECK: [[condition:%.+]] = load i1, ptr [[conditionPtr]]
+// CHECK: [[arg0ptr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 1
+// CHECK: [[arg0:%.+]] = load i32, ptr [[arg0ptr]]
+// CHECK: [[arg1ptr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 2
+// CHECK: [[arg1:%.+]] = load float, ptr [[arg1ptr]]
+// CHECK: [[arg2ptr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 3
+// CHECK: [[arg2:%.+]] = load ptr, ptr [[arg2ptr]]
+// CHECK: [[arg3ptr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 4
+// CHECK: [[arg3:%.+]] = load fp128, ptr [[arg3ptr]]
+// CHECK: [[allocatedPtr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 5
+// CHECK: [[allocated:%.+]] = load ptr, ptr [[allocatedPtr]]
+// CHECK: [[loadedPtr:%.+]] = getelementptr {{.+}}, ptr [[data]], i32 0, i32 6
+// CHECK: [[loaded:%.+]] = load i32, ptr [[loadedPtr]]
+// CHECK: br label
+
+// CHECK: br i1 [[condition]], label %[[true:.+]], label %[[false:.+]]
+// CHECK: [[false]]:
+// CHECK-NEXT: br label
+// CHECK: [[true]]:
+// CHECK: call void @foo(i32 [[arg0]], float [[arg1]], ptr [[arg2]], fp128 [[arg3]], ptr [[allocated]], i32 [[loaded]])
+// CHECK-NEXT: br label
+// CHECK: ret void
+
+// Check the wrapper function
+// CHECK: define void [[wrapperfn]](ptr [[globalTID:%.+]], ptr [[boundTID:%.+]], ptr [[data:%.+]])
+// CHECK: call void [[outlinedfn]](ptr [[data]])
+// CHECK: ret void
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
kiranchandramohan
approved these changes
Oct 3, 2023
| llvm.func @foo() | ||
|
|
||
| // CHECK-LABEL: @omp_teams_simple | ||
| // CHECK: call void {{.*}} @__kmpc_fork_teams(ptr @{{.+}}, i32 0, ptr [[wrapperfn:.+]]) |
There was a problem hiding this comment.
Could you convert the captured variables (eg. wrapperfn) to caps? This is to distinguish them easily from code.
I would like to get it removed, because that is unnecessary (I did not realize this earlier while submitting the patch). But because nobody is actively reviewing #67723, I do not want to delay progress for the construct. So, I will keep updating that PR with updates to testcases while I wait on reviews. The functions should not need any further changes there. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
This patch adds translation from
omp.teamsoperation to LLVM IR using OpenMPIRBuilder. The clauses are not handled in this patch.