Automerge: [mlir][acc] Lower sequential acc.loop to scf.for in ACCComputeLowering (#206165)

Sequential loops already have fixed parallelism, so represent them with
`scf.for` rather than `scf.parallel`. To prevent further analysis and
parallelization, `parDimAttr` is set to seq.

Author: delaram-talaashrafi

mlir/lib/Dialect/OpenACC/Transforms/ACCComputeLowering.cpp

@@ -313,15 +313,14 @@ class ACCLoopConversion : public OpRewritePattern<LoopOp> {
     LoopParMode parMode = loopOp.getDefaultOrDeviceTypeParallelism(deviceType);
 
     if (parMode == LoopParMode::loop_seq || isOpInSerialRegion(loopOp)) {
-      // Although it might seem unintuitive, scf.parallel is used here because
-      // the parallelism of the loop is already predetermined (as sequential).
-      // scf.for will become a candidate for auto-parallelization analysis.
-      auto parallelOp = convertACCLoopToSCFParallel(loopOp, rewriter);
-      if (!parallelOp)
+      // Use scf.for with sequential loops, because the loop's parallelism is
+      // already determined.
+      auto forOp =
+          convertACCLoopToSCFFor(loopOp, rewriter, /*enableCollapse=*/true);
+      if (!forOp)
         return failure();
-      setParDimsAttr(parallelOp,
-                     GPUParallelDimsAttr::seq(loopOp->getContext()));
-      rewriter.replaceOp(loopOp, parallelOp);
+      setParDimsAttr(forOp, GPUParallelDimsAttr::seq(loopOp->getContext()));
+      rewriter.replaceOp(loopOp, forOp);
     } else if (parMode == LoopParMode::loop_auto) {
       // All loops in serial regions should have already been handled.
       assert(!isOpInSerialRegion(loopOp) &&

mlir/test/Dialect/OpenACC/acc-compute-lowering-compute.mlir

@@ -39,7 +39,7 @@ func.func @parallel_seq_loop(%buf: memref<4xi32>) {
   // CHECK: acc.kernel_environment
   // CHECK: acc.par_width {{.*}} {par_dim = #acc.par_dim<block_x>}
   // CHECK: acc.compute_region launch(
-  // CHECK: scf.parallel
+  // CHECK: scf.for
   // CHECK: acc.par_dims = #acc<par_dims[sequential]>
   acc.parallel num_gangs({%c10_i32 : i32}) dataOperands(%dev : memref<4xi32>) {
     acc.loop control(%i : index) = (%c0 : index) to (%c4 : index) step (%c1 : index) {
@@ -66,7 +66,7 @@ func.func @serial_loop(%buf: memref<4xi32>) {
   // CHECK: acc.kernel_environment
   // CHECK: acc.par_width {par_dim = #acc.par_dim<sequential>}
   // CHECK: acc.compute_region launch(
-  // CHECK: scf.parallel
+  // CHECK: scf.for
   // CHECK: acc.par_dims = #acc<par_dims[sequential]>
   acc.serial dataOperands(%dev : memref<4xi32>) {
     acc.loop control(%i : index) = (%c0 : index) to (%c4 : index) step (%c1 : index) {
@@ -150,7 +150,7 @@ func.func @parallel_unit_launch_serial_loops(%buf: memref<4xi32>) {
   // CHECK: acc.kernel_environment
   // CHECK: acc.par_width {par_dim = #acc.par_dim<sequential>}
   // CHECK: acc.compute_region launch(
-  // CHECK: scf.parallel
+  // CHECK: scf.for
   // CHECK: acc.par_dims = #acc<par_dims[sequential]>
   acc.parallel num_gangs({%c1_i32 : i32}) num_workers(%c1_i32 : i32) vector_length(%c1_i32 : i32) dataOperands(%dev : memref<4xi32>) {
     acc.loop control(%i : index) = (%c0 : index) to (%c4 : index) step (%c1 : index) {
@@ -182,7 +182,7 @@ func.func @kernels_unit_launch_serial_loops(%buf: memref<4xi32>) {
   // CHECK: acc.kernel_environment
   // CHECK: acc.par_width {par_dim = #acc.par_dim<sequential>}
   // CHECK: acc.compute_region launch(
-  // CHECK: scf.parallel
+  // CHECK: scf.for
   // CHECK: acc.par_dims = #acc<par_dims[sequential]>
   acc.kernels num_gangs({%c1_i32 : i32}) num_workers(%c1_i32 : i32) vector_length(%c1_i32 : i32) dataOperands(%dev : memref<4xi32>) {
     acc.loop control(%i : index) = (%c0 : index) to (%c4 : index) step (%c1 : index) {

mlir/test/Dialect/OpenACC/acc-compute-lowering-loop.mlir

@@ -95,7 +95,7 @@ func.func @serial_loop_normalized(%buf: memref<1xi32>) {
   // CHECK: acc.kernel_environment
   // CHECK: acc.par_width {par_dim = #acc.par_dim<sequential>}
   // CHECK: acc.compute_region launch(
-  // CHECK: scf.parallel
+  // CHECK: scf.for
   // CHECK-DAG: arith.muli
   // CHECK-DAG: arith.addi
   // CHECK: acc.par_dims = #acc<par_dims[sequential]>
@@ -137,9 +137,9 @@ acc.routine @routine_with_loop func(@device_routine_with_loop) seq
 // CHECK-LABEL: func.func @device_routine_with_loop
 // CHECK: attributes {acc.specialized_routine = #acc.specialized_routine<@routine_with_loop, <seq>, "host_routine_with_loop">}
 // CHECK-NOT: acc.loop
-// CHECK: scf.parallel
+// CHECK: scf.for
 // CHECK: acc.par_dims = #acc<par_dims[sequential]>
-// CHECK-NOT: scf.for
+// CHECK-NOT: scf.parallel
 func.func @device_routine_with_loop(%buf: memref<8xi32>) attributes {acc.specialized_routine = #acc.specialized_routine<@routine_with_loop, <seq>, "host_routine_with_loop">} {
   %c0 = arith.constant 0 : index
   %c1 = arith.constant 1 : index