[OpenMPOpt] Make parallel regions reachable from new DeviceRTL loop functions #150927
Conversation
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
llvmbot
added
llvm:transforms
clang:openmp
|
@llvm/pr-subscribers-offload Author: Sergio Afonso (skatrak) ChangesThis patch updates the OpenMP optimization pass to know about the new DeviceRTL functions for loop constructs. This change marks these functions as potentially containing parallel regions, which fixes a current bug with the state machine rewrite optimization. It previously failed to identify parallel regions located inside of the callbacks passed to these new DeviceRTL functions, causing the resulting code to skip executing these parallel regions. As a result, Generic kernels produced by Flang that contain parallel regions now work properly. One known related issue not fixed by this patch is that the presence of calls to these functions will prevent the SPMD-ization of Generic kernels by OpenMPOpt. Previously, this was due to assuming there was no parallel region. This is changed by this patch, but instead we now mark it temporarily as unsupported in an SPMD context. The reason is that, without additional changes, code intended for the main thread of the team located outside of the parallel region would not be guarded properly, resulting in race conditions and generally invalid behavior. Full diff: https://github.com/llvm/llvm-project/pull/150927.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
index 5e2247f2a88d0..d58da7b1db0e3 100644
--- a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
+++ b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -5020,6 +5020,28 @@ struct AAKernelInfoCallSite : AAKernelInfo {
case OMPRTL___kmpc_free_shared:
// Return without setting a fixpoint, to be resolved in updateImpl.
return;
+ case OMPRTL___kmpc_distribute_static_loop_4:
+ case OMPRTL___kmpc_distribute_static_loop_4u:
+ case OMPRTL___kmpc_distribute_static_loop_8:
+ case OMPRTL___kmpc_distribute_static_loop_8u:
+ case OMPRTL___kmpc_distribute_for_static_loop_4:
+ case OMPRTL___kmpc_distribute_for_static_loop_4u:
+ case OMPRTL___kmpc_distribute_for_static_loop_8:
+ case OMPRTL___kmpc_distribute_for_static_loop_8u:
+ case OMPRTL___kmpc_for_static_loop_4:
+ case OMPRTL___kmpc_for_static_loop_4u:
+ case OMPRTL___kmpc_for_static_loop_8:
+ case OMPRTL___kmpc_for_static_loop_8u:
+ // Parallel regions might be reached by these calls, as they take a
+ // callback argument potentially arbitrary user-provided code.
+ ReachedUnknownParallelRegions.insert(&CB);
+ // TODO: The presence of these calls on their own does not prevent a
+ // kernel from being SPMD-izable. We mark it as such because we need
+ // further changes in order to also consider the contents of the
+ // callbacks passed to them.
+ SPMDCompatibilityTracker.indicatePessimisticFixpoint();
+ SPMDCompatibilityTracker.insert(&CB);
+ break;
default:
// Unknown OpenMP runtime calls cannot be executed in SPMD-mode,
// generally. However, they do not hide parallel regions.
diff --git a/offload/test/offloading/fortran/target-generic-loops.f90 b/offload/test/offloading/fortran/target-generic-loops.f90
new file mode 100644
index 0000000000000..07bcbfd2c8752
--- /dev/null
+++ b/offload/test/offloading/fortran/target-generic-loops.f90
@@ -0,0 +1,130 @@
+! Offloading test for generic target regions containing different kinds of
+! loop constructs inside.
+! REQUIRES: flang, amdgpu
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ integer :: i1, i2, n1, n2, counter
+
+ n1 = 100
+ n2 = 50
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 1 100
+ print '(I2" "I0)', 1, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 2 200
+ print '(I2" "I0)', 2, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp end target
+
+ ! CHECK: 3 203
+ print '(I2" "I0)', 3, counter
+
+ counter = 0
+ !$omp target map(tofrom: counter)
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp end target
+
+ ! CHECK: 4 102
+ print '(I2" "I0)', 4, counter
+
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ end do
+
+ ! CHECK: 5 5000
+ print '(I2" "I0)', 5, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ end do
+
+ ! CHECK: 6 5200
+ print '(I2" "I0)', 6, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ end do
+
+ ! CHECK: 7 10000
+ print '(I2" "I0)', 7, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ end do
+
+ ! CHECK: 8 10300
+ print '(I2" "I0)', 8, counter
+end program
diff --git a/offload/test/offloading/fortran/target-spmd-loops.f90 b/offload/test/offloading/fortran/target-spmd-loops.f90
new file mode 100644
index 0000000000000..7407f0c0768cb
--- /dev/null
+++ b/offload/test/offloading/fortran/target-spmd-loops.f90
@@ -0,0 +1,39 @@
+! Offloading test for generic target regions containing different kinds of
+! loop constructs inside.
+! REQUIRES: flang, amdgpu
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ integer :: i1, n1, counter
+
+ n1 = 100
+
+ counter = 0
+ !$omp target parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+
+ ! CHECK: 1 100
+ print '(I2" "I0)', 1, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 2 100
+ print '(I2" "I0)', 2, counter
+
+ counter = 0
+ !$omp target teams distribute parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+
+ ! CHECK: 3 100
+ print '(I2" "I0)', 3, counter
+end program
|
|
@llvm/pr-subscribers-llvm-transforms Author: Sergio Afonso (skatrak) ChangesThis patch updates the OpenMP optimization pass to know about the new DeviceRTL functions for loop constructs. This change marks these functions as potentially containing parallel regions, which fixes a current bug with the state machine rewrite optimization. It previously failed to identify parallel regions located inside of the callbacks passed to these new DeviceRTL functions, causing the resulting code to skip executing these parallel regions. As a result, Generic kernels produced by Flang that contain parallel regions now work properly. One known related issue not fixed by this patch is that the presence of calls to these functions will prevent the SPMD-ization of Generic kernels by OpenMPOpt. Previously, this was due to assuming there was no parallel region. This is changed by this patch, but instead we now mark it temporarily as unsupported in an SPMD context. The reason is that, without additional changes, code intended for the main thread of the team located outside of the parallel region would not be guarded properly, resulting in race conditions and generally invalid behavior. Full diff: https://github.com/llvm/llvm-project/pull/150927.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
index 5e2247f2a88d0..d58da7b1db0e3 100644
--- a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
+++ b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -5020,6 +5020,28 @@ struct AAKernelInfoCallSite : AAKernelInfo {
case OMPRTL___kmpc_free_shared:
// Return without setting a fixpoint, to be resolved in updateImpl.
return;
+ case OMPRTL___kmpc_distribute_static_loop_4:
+ case OMPRTL___kmpc_distribute_static_loop_4u:
+ case OMPRTL___kmpc_distribute_static_loop_8:
+ case OMPRTL___kmpc_distribute_static_loop_8u:
+ case OMPRTL___kmpc_distribute_for_static_loop_4:
+ case OMPRTL___kmpc_distribute_for_static_loop_4u:
+ case OMPRTL___kmpc_distribute_for_static_loop_8:
+ case OMPRTL___kmpc_distribute_for_static_loop_8u:
+ case OMPRTL___kmpc_for_static_loop_4:
+ case OMPRTL___kmpc_for_static_loop_4u:
+ case OMPRTL___kmpc_for_static_loop_8:
+ case OMPRTL___kmpc_for_static_loop_8u:
+ // Parallel regions might be reached by these calls, as they take a
+ // callback argument potentially arbitrary user-provided code.
+ ReachedUnknownParallelRegions.insert(&CB);
+ // TODO: The presence of these calls on their own does not prevent a
+ // kernel from being SPMD-izable. We mark it as such because we need
+ // further changes in order to also consider the contents of the
+ // callbacks passed to them.
+ SPMDCompatibilityTracker.indicatePessimisticFixpoint();
+ SPMDCompatibilityTracker.insert(&CB);
+ break;
default:
// Unknown OpenMP runtime calls cannot be executed in SPMD-mode,
// generally. However, they do not hide parallel regions.
diff --git a/offload/test/offloading/fortran/target-generic-loops.f90 b/offload/test/offloading/fortran/target-generic-loops.f90
new file mode 100644
index 0000000000000..07bcbfd2c8752
--- /dev/null
+++ b/offload/test/offloading/fortran/target-generic-loops.f90
@@ -0,0 +1,130 @@
+! Offloading test for generic target regions containing different kinds of
+! loop constructs inside.
+! REQUIRES: flang, amdgpu
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ integer :: i1, i2, n1, n2, counter
+
+ n1 = 100
+ n2 = 50
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 1 100
+ print '(I2" "I0)', 1, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 2 200
+ print '(I2" "I0)', 2, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp end target
+
+ ! CHECK: 3 203
+ print '(I2" "I0)', 3, counter
+
+ counter = 0
+ !$omp target map(tofrom: counter)
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp end target
+
+ ! CHECK: 4 102
+ print '(I2" "I0)', 4, counter
+
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ end do
+
+ ! CHECK: 5 5000
+ print '(I2" "I0)', 5, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ end do
+
+ ! CHECK: 6 5200
+ print '(I2" "I0)', 6, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ end do
+
+ ! CHECK: 7 10000
+ print '(I2" "I0)', 7, counter
+
+ counter = 0
+ !$omp target teams distribute reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ !$omp parallel do reduction(+:counter)
+ do i2=1, n2
+ counter = counter + 1
+ end do
+ counter = counter + 1
+ end do
+
+ ! CHECK: 8 10300
+ print '(I2" "I0)', 8, counter
+end program
diff --git a/offload/test/offloading/fortran/target-spmd-loops.f90 b/offload/test/offloading/fortran/target-spmd-loops.f90
new file mode 100644
index 0000000000000..7407f0c0768cb
--- /dev/null
+++ b/offload/test/offloading/fortran/target-spmd-loops.f90
@@ -0,0 +1,39 @@
+! Offloading test for generic target regions containing different kinds of
+! loop constructs inside.
+! REQUIRES: flang, amdgpu
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+program main
+ integer :: i1, n1, counter
+
+ n1 = 100
+
+ counter = 0
+ !$omp target parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+
+ ! CHECK: 1 100
+ print '(I2" "I0)', 1, counter
+
+ counter = 0
+ !$omp target map(tofrom:counter)
+ !$omp parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+ !$omp end target
+
+ ! CHECK: 2 100
+ print '(I2" "I0)', 2, counter
+
+ counter = 0
+ !$omp target teams distribute parallel do reduction(+:counter)
+ do i1=1, n1
+ counter = counter + 1
+ end do
+
+ ! CHECK: 3 100
+ print '(I2" "I0)', 3, counter
+end program
|
This was referenced Jul 28, 2025
skatrak
force-pushed
the
users/skatrak/flang-generic-05-parallel-wrapper
branch
from
8771c0f to
b6e9849
Compare
…unctions This patch updates the OpenMP optimization pass to know about the new DeviceRTL functions for loop constructs. This change marks these functions as potentially containing parallel regions, which fixes a current bug with the state machine rewrite optimization. It previously failed to identify parallel regions located inside of the callbacks passed to these new DeviceRTL functions, causing the resulting code to skip executing these parallel regions. As a result, Generic kernels produced by Flang that contain parallel regions now work properly. One known related issue not fixed by this patch is that the presence of calls to these functions will prevent the SPMD-ization of Generic kernels by OpenMPOpt. Previously, this was due to assuming there was no parallel region. This is changed by this patch, but instead we now mark it temporarily as unsupported in an SPMD context. The reason is that, without additional changes, code intended for the main thread of the team located outside of the parallel region would not be guarded properly, resulting in race conditions and generally invalid behavior.
skatrak
force-pushed
the
users/skatrak/flang-generic-06-openmp-opt
branch
from
aac44b8 to
b6c5ff8
Compare
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 updates the OpenMP optimization pass to know about the new DeviceRTL functions for loop constructs.
This change marks these functions as potentially containing parallel regions, which fixes a current bug with the state machine rewrite optimization. It previously failed to identify parallel regions located inside of the callbacks passed to these new DeviceRTL functions, causing the resulting code to skip executing these parallel regions.
As a result, Generic kernels produced by Flang that contain parallel regions now work properly.
One known related issue not fixed by this patch is that the presence of calls to these functions will prevent the SPMD-ization of Generic kernels by OpenMPOpt. Previously, this was due to assuming there was no parallel region. This is changed by this patch, but instead we now mark it temporarily as unsupported in an SPMD context. The reason is that, without additional changes, code intended for the main thread of the team located outside of the parallel region would not be guarded properly, resulting in race conditions and generally invalid behavior.