[llvm-branch-commits] [mlir] [MLIR][OpenMP] Make omp.distribute into a loop wrapper (PR #87239)
Sergio Afonso via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Tue Apr 2 07:50:37 PDT 2024
https://github.com/skatrak updated https://github.com/llvm/llvm-project/pull/87239
>From 4537071171506b17de3727800e3754e412c9a967 Mon Sep 17 00:00:00 2001
From: Sergio Afonso <safonsof at amd.com>
Date: Mon, 1 Apr 2024 14:08:33 +0100
Subject: [PATCH] [MLIR][OpenMP] Make omp.distribute into a loop wrapper
This patch updates the definition of `omp.distribute` to enforce the
restrictions of a wrapper operation.
---
mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td | 22 +++++++++--
mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp | 11 ++++++
mlir/test/Dialect/OpenMP/invalid.mlir | 34 ++++++++++++++++-
mlir/test/Dialect/OpenMP/ops.mlir | 38 +++++++++++++++----
4 files changed, 93 insertions(+), 12 deletions(-)
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
index a7bf93deae2fb3..8dbfe447616e11 100644
--- a/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
+++ b/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
@@ -816,7 +816,8 @@ def YieldOp : OpenMP_Op<"yield",
//===----------------------------------------------------------------------===//
def DistributeOp : OpenMP_Op<"distribute", [AttrSizedOperandSegments,
DeclareOpInterfaceMethods<LoopWrapperInterface>,
- RecursiveMemoryEffects]> {
+ RecursiveMemoryEffects,
+ SingleBlockImplicitTerminator<"TerminatorOp">]> {
let summary = "distribute construct";
let description = [{
The distribute construct specifies that the iterations of one or more loops
@@ -831,15 +832,28 @@ def DistributeOp : OpenMP_Op<"distribute", [AttrSizedOperandSegments,
The distribute loop construct specifies that the iterations of the loop(s)
will be executed in parallel by threads in the current context. These
iterations are spread across threads that already exist in the enclosing
- region. The lower and upper bounds specify a half-open range: the
- range includes the lower bound but does not include the upper bound. If the
- `inclusive` attribute is specified then the upper bound is also included.
+ region.
+
+ The body region can contain a single block which must contain a single
+ operation and a terminator. The operation must be another compatible loop
+ wrapper or an `omp.loop_nest`.
The `dist_schedule_static` attribute specifies the schedule for this
loop, determining how the loop is distributed across the parallel threads.
The optional `schedule_chunk` associated with this determines further
controls this distribution.
+ ```mlir
+ omp.distribute <clauses> {
+ omp.loop_nest (%i1, %i2) : index = (%c0, %c0) to (%c10, %c10) step (%c1, %c1) {
+ %a = load %arrA[%i1, %i2] : memref<?x?xf32>
+ %b = load %arrB[%i1, %i2] : memref<?x?xf32>
+ %sum = arith.addf %a, %b : f32
+ store %sum, %arrC[%i1, %i2] : memref<?x?xf32>
+ omp.yield
+ }
+ }
+ ```
// TODO: private_var, firstprivate_var, lastprivate_var, collapse
}];
let arguments = (ins
diff --git a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
index 564c23201db4fd..b407d27ef53e39 100644
--- a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
+++ b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
@@ -1516,6 +1516,17 @@ LogicalResult DistributeOp::verify() {
return emitError(
"expected equal sizes for allocate and allocator variables");
+ if (!isWrapper())
+ return emitOpError() << "must be a loop wrapper";
+
+ if (LoopWrapperInterface nested = getNestedWrapper()) {
+ // Check for the allowed leaf constructs that may appear in a composite
+ // construct directly after DISTRIBUTE.
+ if (!isa<ParallelOp, SimdLoopOp>(nested))
+ return emitError() << "only supported nested wrappers are 'omp.parallel' "
+ "and 'omp.simdloop'";
+ }
+
return success();
}
diff --git a/mlir/test/Dialect/OpenMP/invalid.mlir b/mlir/test/Dialect/OpenMP/invalid.mlir
index 8f4103dabee5df..35f5d24deb5d17 100644
--- a/mlir/test/Dialect/OpenMP/invalid.mlir
+++ b/mlir/test/Dialect/OpenMP/invalid.mlir
@@ -1847,7 +1847,16 @@ func.func @omp_target_depend(%data_var: memref<i32>) {
// -----
-func.func @omp_distribute(%data_var : memref<i32>) -> () {
+func.func @omp_distribute_schedule(%chunk_size : i32) -> () {
+ // expected-error @below {{op chunk size set without dist_schedule_static being present}}
+ "omp.distribute"(%chunk_size) <{operandSegmentSizes = array<i32: 1, 0, 0>}> ({
+ "omp.terminator"() : () -> ()
+ }) : (i32) -> ()
+}
+
+// -----
+
+func.func @omp_distribute_allocate(%data_var : memref<i32>) -> () {
// expected-error @below {{expected equal sizes for allocate and allocator variables}}
"omp.distribute"(%data_var) <{operandSegmentSizes = array<i32: 0, 1, 0>}> ({
"omp.terminator"() : () -> ()
@@ -1856,6 +1865,29 @@ func.func @omp_distribute(%data_var : memref<i32>) -> () {
// -----
+func.func @omp_distribute_wrapper() -> () {
+ // expected-error @below {{op must be a loop wrapper}}
+ "omp.distribute"() ({
+ %0 = arith.constant 0 : i32
+ "omp.terminator"() : () -> ()
+ }) : () -> ()
+}
+
+// -----
+
+func.func @omp_distribute_nested_wrapper(%data_var : memref<i32>) -> () {
+ // expected-error @below {{only supported nested wrappers are 'omp.parallel' and 'omp.simdloop'}}
+ "omp.distribute"() ({
+ "omp.wsloop"() ({
+ %0 = arith.constant 0 : i32
+ "omp.terminator"() : () -> ()
+ }) : () -> ()
+ "omp.terminator"() : () -> ()
+ }) : () -> ()
+}
+
+// -----
+
omp.private {type = private} @x.privatizer : i32 alloc {
^bb0(%arg0: i32):
%0 = arith.constant 0.0 : f32
diff --git a/mlir/test/Dialect/OpenMP/ops.mlir b/mlir/test/Dialect/OpenMP/ops.mlir
index 8d9acab67e0358..a7b0832eff21f3 100644
--- a/mlir/test/Dialect/OpenMP/ops.mlir
+++ b/mlir/test/Dialect/OpenMP/ops.mlir
@@ -559,30 +559,54 @@ func.func @omp_simdloop_pretty_multiple(%lb1 : index, %ub1 : index, %step1 : ind
}
// CHECK-LABEL: omp_distribute
-func.func @omp_distribute(%chunk_size : i32, %data_var : memref<i32>) -> () {
+func.func @omp_distribute(%chunk_size : i32, %data_var : memref<i32>, %arg0 : i32) -> () {
// CHECK: omp.distribute
"omp.distribute" () ({
- omp.terminator
+ "omp.loop_nest" (%arg0, %arg0, %arg0) ({
+ ^bb0(%iv: i32):
+ "omp.yield"() : () -> ()
+ }) : (i32, i32, i32) -> ()
+ "omp.terminator"() : () -> ()
}) {} : () -> ()
// CHECK: omp.distribute
omp.distribute {
- omp.terminator
+ omp.loop_nest (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
}
// CHECK: omp.distribute dist_schedule_static
omp.distribute dist_schedule_static {
- omp.terminator
+ omp.loop_nest (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
}
// CHECK: omp.distribute dist_schedule_static chunk_size(%{{.+}} : i32)
omp.distribute dist_schedule_static chunk_size(%chunk_size : i32) {
- omp.terminator
+ omp.loop_nest (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
}
// CHECK: omp.distribute order(concurrent)
omp.distribute order(concurrent) {
- omp.terminator
+ omp.loop_nest (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
}
// CHECK: omp.distribute allocate(%{{.+}} : memref<i32> -> %{{.+}} : memref<i32>)
omp.distribute allocate(%data_var : memref<i32> -> %data_var : memref<i32>) {
- omp.terminator
+ omp.loop_nest (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
+ }
+ // CHECK: omp.distribute
+ omp.distribute {
+ // TODO Remove induction variables from omp.simdloop.
+ omp.simdloop for (%iv) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.loop_nest (%iv2) : i32 = (%arg0) to (%arg0) step (%arg0) {
+ omp.yield
+ }
+ omp.yield
+ }
}
return
}
More information about the llvm-branch-commits
mailing list