[Mlir-commits] [mlir] [openacc] Attach Parallelism Levels to Auto Loops (PR #200884)

[Delaram Talaashrafi]

https://github.com/delaram-talaashrafi created https://github.com/llvm/llvm-project/pull/200884

Auto loops are analyzed by the compiler in later compilation stages to determine whether they can be parallelized. These loops may carry parallelism levels (this does not guarantee that they are parallelizable, compiler should still analyze them). However, if the loop is parallelized, the parallelism levels specified in the source should be respected. This change attaches the parallelism level to auto loops, which enables their propagation through next compilation steps.

>From 7df55e9452c631002a16afee8edb427af252ef46 Mon Sep 17 00:00:00 2001
From: Delaram Talaashrafi <dtalaashrafi at rome5.pgi.net>
Date: Mon, 1 Jun 2026 10:38:20 -0700
Subject: [PATCH] [openacc] Attach Parallelism Levels to Auto Loops

Auto loops are analyzed by the compiler in later compilation stages to determine
whether they can be parallelized. These loops may carry a parallelism level, but
this does not guarantee that they are parallelizable (the compiler should still
analyze them). However, if a loop is parallelized, the parallelism level specified
in the source should be respected. This change attaches the parallelism level to
auto loops and enables its propagation through subsequent compilation steps.
---
 .../OpenACC/Transforms/ACCComputeLowering.cpp |  7 +++++
 .../OpenACC/acc-compute-lowering-loop.mlir    | 29 +++++++++++++++++++
 2 files changed, 36 insertions(+)

diff --git a/mlir/lib/Dialect/OpenACC/Transforms/ACCComputeLowering.cpp b/mlir/lib/Dialect/OpenACC/Transforms/ACCComputeLowering.cpp
index 9868584a4b699..80b4570587a17 100644
--- a/mlir/lib/Dialect/OpenACC/Transforms/ACCComputeLowering.cpp
+++ b/mlir/lib/Dialect/OpenACC/Transforms/ACCComputeLowering.cpp
@@ -314,6 +314,13 @@ class ACCLoopConversion : public OpRewritePattern<LoopOp> {
           convertACCLoopToSCFFor(loopOp, rewriter, /*enableCollapse=*/true);
       if (!forOp)
         return failure();
+      SmallVector<GPUParallelDimAttr> parDims =
+          getParallelDimensions(loopOp, policy, deviceType);
+      if (!parDims.empty()) {
+        auto parDimsAttr =
+            GPUParallelDimsAttr::get(loopOp->getContext(), parDims);
+        setParDimsAttr(forOp, parDimsAttr);
+      }
       rewriter.replaceOp(loopOp, forOp);
     } else if (!isOpInComputeRegion(loopOp) &&
                !isSpecializedAccRoutine(
diff --git a/mlir/test/Dialect/OpenACC/acc-compute-lowering-loop.mlir b/mlir/test/Dialect/OpenACC/acc-compute-lowering-loop.mlir
index 4032fed217b62..358d1328fa028 100644
--- a/mlir/test/Dialect/OpenACC/acc-compute-lowering-loop.mlir
+++ b/mlir/test/Dialect/OpenACC/acc-compute-lowering-loop.mlir
@@ -174,3 +174,32 @@ func.func @device_routine_vector_with_loop(%buf: memref<8xi32>) attributes {acc.
   } attributes {independent = [#acc.device_type<none>], vector = [#acc.device_type<none>]}
   return
 }
+
+// -----
+
+// Auto loop with gang: lowered to scf.for with predetermined par_dims.
+// CHECK-LABEL: func.func @parallel_loop_auto_gang
+func.func @parallel_loop_auto_gang(%buf: memref<1xi32>) {
+  %c0 = arith.constant 0 : index
+  %c1_i32 = arith.constant 1 : i32
+  %c10_i32 = arith.constant 10 : i32
+  %c100_i32 = arith.constant 100 : i32
+
+  %dev = acc.copyin varPtr(%buf : memref<1xi32>) -> memref<1xi32>
+  // CHECK-NOT: acc.parallel
+  // CHECK: acc.kernel_environment
+  // CHECK: acc.par_width {{.*}} {par_dim = #acc.par_dim<block_x>}
+  // CHECK: acc.compute_region launch(
+  // CHECK: scf.for
+  // CHECK-NOT: scf.parallel
+  // CHECK: acc.par_dims = #acc<par_dims[block_x]>
+  acc.parallel num_gangs({%c10_i32 : i32}) dataOperands(%dev : memref<1xi32>) {
+    acc.loop gang control(%arg0 : i32) = (%c1_i32 : i32) to (%c100_i32 : i32) step (%c1_i32 : i32) {
+      memref.store %arg0, %dev[%c0] : memref<1xi32>
+      acc.yield
+    } attributes {auto_ = [#acc.device_type<none>]}
+    acc.yield
+  }
+  acc.copyout accPtr(%dev : memref<1xi32>) to varPtr(%buf : memref<1xi32>)
+  return
+}