Purge scf.parallel op containing only wait_all no-ops (Xilinx#569)

nod-ai · May 9, 2024 · 85a99b3 · 85a99b3
1 parent ccca6b5
commit 85a99b3
Show file tree

Hide file tree

Showing 2 changed files with 66 additions and 0 deletions.
diff --git a/mlir/lib/Conversion/AIRRtToNpuPass.cpp b/mlir/lib/Conversion/AIRRtToNpuPass.cpp
@@ -871,6 +871,7 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
     moveFuncOpToEndOfDeviceOp(module);
 
     // Purge all wait all ops
+    purgeSCFParContainingOnlyWaitAllOps(module);
     purgeWaitAlls(module);
 
     // Purge airrt.dma x and y fields, as they are obsolete for AIE2.
@@ -1131,6 +1132,38 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
     }
   }
 
+  void purgeSCFParContainingOnlyWaitAllOps(ModuleOp module) {
+    SmallVector<scf::ParallelOp> scf_pars;
+    module.walk([&](mlir::func::FuncOp f) {
+      f.walk([&](scf::ParallelOp par_op) { scf_pars.push_back(par_op); });
+    });
+    OpBuilder builder(module);
+    for (auto par_op : scf_pars) {
+      bool containsOnlyWaitAll = true;
+      par_op.walk([&](Operation *o) {
+        if (isa<airrt::WaitAllOp>(o))
+          return;
+        else if (isa<scf::ParallelOp>(o))
+          return;
+        else if (o->mightHaveTrait<OpTrait::IsTerminator>())
+          return;
+        else {
+          containsOnlyWaitAll = false;
+          return;
+        }
+      });
+      if (!containsOnlyWaitAll)
+        assert(false && "found scf.parallel op at this IR, NYI");
+      builder.setInsertionPoint(par_op);
+      auto newWaitAll = builder.create<airrt::WaitAllOp>(
+          par_op->getLoc(), airrt::EventType::get(par_op->getContext()),
+          par_op.getInitVals());
+      for (auto res : par_op->getResults())
+        res.replaceAllUsesWith(newWaitAll->getResult(0));
+      par_op->erase();
+    }
+  }
+
   std::optional<AIE::ShimDMAAllocationOp>
   getAllocOpForSymbol(SmallVector<AIE::ShimDMAAllocationOp> shimDmaAllocOps,
                       StringRef sym_name) {

diff --git a/mlir/test/Conversion/AIRRtToNpu/airrt_to_npu.mlir b/mlir/test/Conversion/AIRRtToNpu/airrt_to_npu.mlir
@@ -897,3 +897,36 @@ module {
     return
   }
 }
+
+// -----
+
+// Purge scf.parallel op which contains only no-ops.
+
+// CHECK-LABEL: func20
+// CHECK: return
+module {
+  func.func @func20() {
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %c2 = arith.constant 2 : index
+    %c152 = arith.constant 152 : index
+    %51 = airrt.wait_all : !airrt.event
+    %52 = scf.for %arg3 = %c0 to %c152 step %c1 iter_args(%arg4 = %51) -> (!airrt.event) {
+      %61 = airrt.wait_all : !airrt.event
+      %62 = airrt.wait_all %arg4, %61 : !airrt.event
+      %63 = airrt.wait_all : !airrt.event
+      %64 = airrt.wait_all %arg4, %63 : !airrt.event
+      %65 = scf.parallel (%arg5) = (%c0) to (%c2) step (%c1) init (%arg4) -> !airrt.event {
+        %66 = airrt.wait_all : !airrt.event
+        %67 = airrt.wait_all %arg4, %66 : !airrt.event
+        scf.reduce(%67 : !airrt.event) {
+        ^bb0(%arg6: !airrt.event, %arg7: !airrt.event):
+          %68 = airrt.wait_all %arg6, %arg7 : !airrt.event
+          scf.reduce.return %68 : !airrt.event
+        }
+      }
+      scf.yield %65 : !airrt.event
+    }
+    return
+  }
+}