[Mlir-commits] [mlir] 9006bc8 - [OpenMP] Enable simd in non-reduction composite constructs (#146097)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Tue Jul 8 06:27:37 PDT 2025
Author: Kajetan Puchalski
Date: 2025-07-08T14:27:33+01:00
New Revision: 9006bc87170e347a776f34439fe1884d719e0861
URL: https://github.com/llvm/llvm-project/commit/9006bc87170e347a776f34439fe1884d719e0861
DIFF: https://github.com/llvm/llvm-project/commit/9006bc87170e347a776f34439fe1884d719e0861.diff
LOG: [OpenMP] Enable simd in non-reduction composite constructs (#146097)
Despite currently being ignored with a warning, simd as a leaf in
composite constructs behaves as expected when the construct does not
contain a reduction. Enable it for those non-reduction constructs.
---------
Signed-off-by: Kajetan Puchalski <kajetan.puchalski at arm.com>
Added:
mlir/test/Target/LLVMIR/openmp-parallel-do-simd.mlir
mlir/test/Target/LLVMIR/openmp-teams-distribute-parallel-do-simd.mlir
Modified:
mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
mlir/test/Target/LLVMIR/openmp-todo.mlir
Removed:
################################################################################
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 7a517fb86d1ff..4ea96703ec0df 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2852,11 +2852,10 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
auto simdOp = cast<omp::SimdOp>(opInst);
- // TODO: Replace this with proper composite translation support.
- // Currently, simd information on composite constructs is ignored, so e.g.
- // 'do/for simd' will be treated the same as a standalone 'do/for'. This is
- // allowed by the spec, since it's equivalent to using a SIMD length of 1.
- if (simdOp.isComposite()) {
+ // Ignore simd in composite constructs with unsupported clauses
+ // TODO: Replace this once simd + clause combinations are properly supported
+ if (simdOp.isComposite() &&
+ (simdOp.getReductionByref().has_value() || simdOp.getIfExpr())) {
if (failed(convertIgnoredWrapper(simdOp, moduleTranslation)))
return failure();
diff --git a/mlir/test/Target/LLVMIR/openmp-parallel-do-simd.mlir b/mlir/test/Target/LLVMIR/openmp-parallel-do-simd.mlir
new file mode 100644
index 0000000000000..bfa57b85245ca
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-parallel-do-simd.mlir
@@ -0,0 +1,38 @@
+// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s
+
+// Check that omp.simd as a leaf of a composite construct still generates
+// the appropriate loop vectorization attribute.
+
+// CHECK-LABEL: define internal void @test_parallel_do_simd..omp_par
+// CHECK: omp.par.entry:
+// CHECK: omp.par.region:
+// CHECK: omp_loop.header:
+// CHECK: omp_loop.inc:
+// CHECK-NEXT: %omp_loop.next = add nuw i32 %omp_loop.iv, 1
+// CHECK-NEXT: br label %omp_loop.header, !llvm.loop ![[LOOP_ATTR:.*]]
+// CHECK: ![[LOOP_ATTR]] = distinct !{![[LOOP_ATTR]], ![[LPAR:.*]], ![[LVEC:.*]]}
+// CHECK: ![[LPAR]] = !{!"llvm.loop.parallel_accesses", ![[PAR_ACC:.*]]}
+// CHECK: ![[LVEC]] = !{!"llvm.loop.vectorize.enable", i1 true}
+
+llvm.func @test_parallel_do_simd() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 {bindc_name = "i"} : (i64) -> !llvm.ptr
+ %2 = llvm.mlir.constant(1000 : i32) : i32
+ %3 = llvm.mlir.constant(1 : i32) : i32
+ %4 = llvm.mlir.constant(1 : i64) : i64
+ omp.parallel {
+ %5 = llvm.mlir.constant(1 : i64) : i64
+ %6 = llvm.alloca %5 x i32 {bindc_name = "i", pinned} : (i64) -> !llvm.ptr
+ %7 = llvm.mlir.constant(1 : i64) : i64
+ omp.wsloop {
+ omp.simd {
+ omp.loop_nest (%arg0) : i32 = (%3) to (%2) inclusive step (%3) {
+ llvm.store %arg0, %6 : i32, !llvm.ptr
+ omp.yield
+ }
+ } {omp.composite}
+ } {omp.composite}
+ omp.terminator
+ }
+ llvm.return
+}
diff --git a/mlir/test/Target/LLVMIR/openmp-teams-distribute-parallel-do-simd.mlir b/mlir/test/Target/LLVMIR/openmp-teams-distribute-parallel-do-simd.mlir
new file mode 100644
index 0000000000000..4d766cc1ac4f4
--- /dev/null
+++ b/mlir/test/Target/LLVMIR/openmp-teams-distribute-parallel-do-simd.mlir
@@ -0,0 +1,51 @@
+// RUN: mlir-translate --mlir-to-llvmir %s | FileCheck %s
+
+// Check that omp.simd as a leaf of a composite construct still generates
+// the appropriate loop vectorization attribute.
+
+// CHECK-LABEL: define internal void @test_teams_distribute_parallel_do_simd..omp_par.2
+// CHECK: teams.body:
+// CHECK: omp.teams.region:
+
+// CHECK-LABEL: define internal void @test_teams_distribute_parallel_do_simd..omp_par.1
+// CHECK: omp.par.entry:
+// CHECK: omp.par.region:
+// CHECK: distribute.exit:
+
+// CHECK-LABEL: define internal void @test_teams_distribute_parallel_do_simd..omp_par
+// CHECK: distribute.body:
+// CHECK: omp.distribute.region:
+// CHECK: omp_loop.header:
+// CHECK: omp_loop.inc:
+// CHECK-NEXT: %omp_loop.next = add nuw i32 %omp_loop.iv, 1
+// CHECK-NEXT: br label %omp_loop.header, !llvm.loop ![[LOOP_ATTR:.*]]
+
+// CHECK: ![[LOOP_ATTR]] = distinct !{![[LOOP_ATTR]], ![[LPAR:.*]], ![[LVEC:.*]]}
+// CHECK: ![[LPAR]] = !{!"llvm.loop.parallel_accesses", ![[PAR_ACC:.*]]}
+// CHECK: ![[LVEC]] = !{!"llvm.loop.vectorize.enable", i1 true}
+
+omp.private {type = private} @_QFEi_private_i32 : i32
+llvm.func @test_teams_distribute_parallel_do_simd() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 {bindc_name = "i"} : (i64) -> !llvm.ptr
+ %2 = llvm.mlir.constant(1000 : i32) : i32
+ %3 = llvm.mlir.constant(1 : i32) : i32
+ %4 = llvm.mlir.constant(1 : i64) : i64
+ omp.teams {
+ omp.parallel {
+ omp.distribute {
+ omp.wsloop {
+ omp.simd private(@_QFEi_private_i32 %1 -> %arg0 : !llvm.ptr) {
+ omp.loop_nest (%arg1) : i32 = (%3) to (%2) inclusive step (%3) {
+ llvm.store %arg1, %arg0 : i32, !llvm.ptr
+ omp.yield
+ }
+ } {omp.composite}
+ } {omp.composite}
+ } {omp.composite}
+ omp.terminator
+ } {omp.composite}
+ omp.terminator
+ }
+ llvm.return
+}
diff --git a/mlir/test/Target/LLVMIR/openmp-todo.mlir b/mlir/test/Target/LLVMIR/openmp-todo.mlir
index 29725a02c075a..e2c32b254c200 100644
--- a/mlir/test/Target/LLVMIR/openmp-todo.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-todo.mlir
@@ -26,20 +26,6 @@ llvm.func @atomic_hint(%v : !llvm.ptr, %x : !llvm.ptr, %expr : i32) {
// -----
-llvm.func @do_simd(%lb : i32, %ub : i32, %step : i32) {
- omp.wsloop {
- // expected-warning at below {{simd information on composite construct discarded}}
- omp.simd {
- omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
- omp.yield
- }
- } {omp.composite}
- } {omp.composite}
- llvm.return
-}
-
-// -----
-
llvm.func @distribute_allocate(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
// expected-error at below {{not yet implemented: Unhandled clause allocate in omp.distribute operation}}
// expected-error at below {{LLVM Translation failed for operation: omp.distribute}}
@@ -503,3 +489,43 @@ llvm.func @wsloop_order(%lb : i32, %ub : i32, %step : i32) {
}
llvm.return
}
+
+// -----
+
+llvm.func @do_simd_if(%1 : !llvm.ptr, %5 : i32, %4 : i32, %6 : i1) {
+ omp.wsloop {
+ // expected-warning at below {{simd information on composite construct discarded}}
+ omp.simd if(%6) {
+ omp.loop_nest (%arg0) : i32 = (%5) to (%4) inclusive step (%5) {
+ llvm.store %arg0, %1 : i32, !llvm.ptr
+ omp.yield
+ }
+ } {omp.composite}
+ } {omp.composite}
+ llvm.return
+}
+
+// -----
+
+omp.declare_reduction @add_reduction_i32 : i32 init {
+^bb0(%arg0: i32):
+ %0 = llvm.mlir.constant(0 : i32) : i32
+ omp.yield(%0 : i32)
+} combiner {
+^bb0(%arg0: i32, %arg1: i32):
+ %0 = llvm.add %arg0, %arg1 : i32
+ omp.yield(%0 : i32)
+}
+llvm.func @do_simd_reduction(%1 : !llvm.ptr, %3 : !llvm.ptr, %6 : i32, %7 : i32) {
+ omp.wsloop reduction(@add_reduction_i32 %3 -> %arg0 : !llvm.ptr) {
+ // expected-warning at below {{simd information on composite construct discarded}}
+ omp.simd reduction(@add_reduction_i32 %arg0 -> %arg1 : !llvm.ptr) {
+ omp.loop_nest (%arg2) : i32 = (%7) to (%6) inclusive step (%7) {
+ llvm.store %arg2, %1 : i32, !llvm.ptr
+ %12 = llvm.load %arg1 : !llvm.ptr -> i32
+ omp.yield
+ }
+ } {omp.composite}
+ } {omp.composite}
+ llvm.return
+}
More information about the Mlir-commits
mailing list