[MLIR][Vector] Add warp distribution for scf.if

[akroviakov]

This PR adds scf.if op distribution to the existing VectorDistribute patterns.
The logic mostly follows that of scf.for: move op outside, wrap each branch with gpu.warp_execute_on_lane_0.
A notable difference to scf.for is that each branch has its own set of escaping values, and scf.if itself does not have block arguments.

[llvmbot]

@llvm/pr-subscribers-mlir-gpu
@llvm/pr-subscribers-mlir-vector

@llvm/pr-subscribers-mlir

Author: Artem Kroviakov (akroviakov)

Changes

This PR adds scf.if op distribution to the existing VectorDistribute patterns.
The logic mostly follows that of scf.for: move op outside, wrap each branch with gpu.warp_execute_on_lane_0.
A notable difference to scf.for is that each branch has its own set of escaping values, and scf.if itself does not have block arguments.

---

Full diff: https://github.com/llvm/llvm-project/pull/157119.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp (+201)
• (modified) mlir/test/Dialect/Vector/vector-warp-distribute.mlir (+69)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
index c84eb2c9f8857..cf5928278aa64 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
@@ -1713,6 +1713,205 @@ struct WarpOpInsert : public WarpDistributionPattern {
   }
 };
 
+struct WarpOpScfIfOp : public WarpDistributionPattern {
+  WarpOpScfIfOp(MLIRContext *ctx, DistributionMapFn fn, PatternBenefit b = 1)
+      : WarpDistributionPattern(ctx, b), distributionMapFn(std::move(fn)) {}
+  LogicalResult matchAndRewrite(WarpExecuteOnLane0Op warpOp,
+                                PatternRewriter &rewriter) const override {
+    gpu::YieldOp warpOpYield = warpOp.getTerminator();
+    // Only pick up `IfOp` if it is the last op in the region.
+    Operation *lastNode = warpOpYield->getPrevNode();
+    auto ifOp = dyn_cast_or_null<scf::IfOp>(lastNode);
+    if (!ifOp)
+      return failure();
+
+    // The current `WarpOp` can yield two types of values:
+    // 1. Not results of `IfOp`:
+    //     Preserve them in the new `WarpOp`.
+    //     Collect their yield index.
+    // 2. Results of `IfOp`:
+    //     They are not part of the new `WarpOp` results.
+    //     Map current warp's yield operand index to `IfOp` result idx.
+    SmallVector<Value> nonIfYieldValues;
+    SmallVector<unsigned> nonIfYieldIndices;
+    llvm::SmallDenseMap<unsigned, unsigned> ifResultMapping;
+    llvm::SmallDenseMap<unsigned, VectorType> ifResultDistTypes;
+    for (OpOperand &yieldOperand : warpOpYield->getOpOperands()) {
+      const unsigned yieldOperandIdx = yieldOperand.getOperandNumber();
+      if (yieldOperand.get().getDefiningOp() != ifOp.getOperation()) {
+        nonIfYieldValues.push_back(yieldOperand.get());
+        nonIfYieldIndices.push_back(yieldOperandIdx);
+        continue;
+      }
+      OpResult ifResult = cast<OpResult>(yieldOperand.get());
+      const unsigned ifResultIdx = ifResult.getResultNumber();
+      ifResultMapping[yieldOperandIdx] = ifResultIdx;
+      // If this `ifOp` result is vector type and it is yielded by the
+      // `WarpOp`, we keep track the distributed type for this result.
+      if (!isa<VectorType>(ifResult.getType()))
+        continue;
+      VectorType distType =
+          cast<VectorType>(warpOp.getResult(yieldOperandIdx).getType());
+      ifResultDistTypes[ifResultIdx] = distType;
+    }
+
+    // Collect `WarpOp`-defined values used in `ifOp`, the new warp op returns
+    // them
+    auto getEscapingValues = [&](Region &branch,
+                                 llvm::SmallSetVector<Value, 32> &values,
+                                 SmallVector<Type> &inputTypes,
+                                 SmallVector<Type> &distTypes) {
+      if (branch.empty())
+        return;
+      mlir::visitUsedValuesDefinedAbove(branch, [&](OpOperand *operand) {
+        Operation *parent = operand->get().getParentRegion()->getParentOp();
+        if (warpOp->isAncestor(parent)) {
+          if (!values.insert(operand->get()))
+            return;
+          Type distType = operand->get().getType();
+          if (auto vecType = dyn_cast<VectorType>(distType)) {
+            AffineMap map = distributionMapFn(operand->get());
+            distType = getDistributedType(vecType, map, warpOp.getWarpSize());
+          }
+          inputTypes.push_back(operand->get().getType());
+          distTypes.push_back(distType);
+        }
+      });
+    };
+    llvm::SmallSetVector<Value, 32> escapingValuesThen;
+    SmallVector<Type> escapingValueInputTypesThen; // inner warp op block args
+    SmallVector<Type> escapingValueDistTypesThen;  // new warp returns
+    getEscapingValues(ifOp.getThenRegion(), escapingValuesThen,
+                      escapingValueInputTypesThen, escapingValueDistTypesThen);
+    llvm::SmallSetVector<Value, 32> escapingValuesElse;
+    SmallVector<Type> escapingValueInputTypesElse; // inner warp op block args
+    SmallVector<Type> escapingValueDistTypesElse;  // new warp returns
+    getEscapingValues(ifOp.getElseRegion(), escapingValuesElse,
+                      escapingValueInputTypesElse, escapingValueDistTypesElse);
+
+    if (llvm::is_contained(escapingValueDistTypesThen, Type{}) ||
+        llvm::is_contained(escapingValueDistTypesElse, Type{}))
+      return failure();
+
+    // The new `WarpOp` groups yields values in following order:
+    // 1. Escaping values then branch
+    // 2. Escaping values else branch
+    // 3. All non-`ifOp` yielded values.
+    SmallVector<Value> newWarpOpYieldValues{escapingValuesThen.begin(),
+                                            escapingValuesThen.end()};
+    newWarpOpYieldValues.append(escapingValuesElse.begin(),
+                                escapingValuesElse.end());
+    SmallVector<Type> newWarpOpDistTypes = escapingValueDistTypesThen;
+    newWarpOpDistTypes.append(escapingValueDistTypesElse.begin(),
+                              escapingValueDistTypesElse.end());
+
+    llvm::SmallDenseMap<unsigned, unsigned> origToNewYieldIdx;
+    for (auto [idx, val] :
+         llvm::zip_equal(nonIfYieldIndices, nonIfYieldValues)) {
+      origToNewYieldIdx[idx] = newWarpOpYieldValues.size();
+      newWarpOpYieldValues.push_back(val);
+      newWarpOpDistTypes.push_back(warpOp.getResult(idx).getType());
+    }
+    // Create the new `WarpOp` with the updated yield values and types.
+    WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndReplaceReturns(
+        rewriter, warpOp, newWarpOpYieldValues, newWarpOpDistTypes);
+
+    // `ifOp` returns the result of the inner warp op.
+    SmallVector<Type> newIfOpDistResTypes;
+    for (auto [i, res] : llvm::enumerate(ifOp.getResults())) {
+      Type distType = cast<Value>(res).getType();
+      if (auto vecType = dyn_cast<VectorType>(distType)) {
+        AffineMap map = distributionMapFn(cast<Value>(res));
+        distType = ifResultDistTypes.count(i)
+                       ? ifResultDistTypes[i]
+                       : getDistributedType(vecType, map, warpOp.getWarpSize());
+      }
+      newIfOpDistResTypes.push_back(distType);
+    }
+    // Create a new `IfOp` outside the new `WarpOp` region.
+    OpBuilder::InsertionGuard g(rewriter);
+    rewriter.setInsertionPointAfter(newWarpOp);
+    auto newIfOp = scf::IfOp::create(rewriter, ifOp.getLoc(),
+                                     newIfOpDistResTypes, ifOp.getCondition(),
+                                     static_cast<bool>(ifOp.thenBlock()),
+                                     static_cast<bool>(ifOp.elseBlock()));
+
+    auto processBranch = [&](Block *oldIfBranch, Block *newIfBranch,
+                             llvm::SmallSetVector<Value, 32> &escapingValues,
+                             SmallVector<Type> &escapingValueInputTypes) {
+      OpBuilder::InsertionGuard g(rewriter);
+      if (!newIfBranch)
+        return;
+      rewriter.setInsertionPointToStart(newIfBranch);
+      llvm::SmallDenseMap<Value, int64_t> escapeValToBlockArgIndex;
+      SmallVector<Value> innerWarpInputVals;
+      SmallVector<Type> innerWarpInputTypes;
+      for (size_t i = 0; i < escapingValues.size(); ++i) {
+        innerWarpInputVals.push_back(newWarpOp.getResult(i));
+        escapeValToBlockArgIndex[escapingValues[i]] =
+            innerWarpInputTypes.size();
+        innerWarpInputTypes.push_back(escapingValueInputTypes[i]);
+      }
+      auto innerWarp = WarpExecuteOnLane0Op::create(
+          rewriter, newWarpOp.getLoc(), newIfOp.getResultTypes(),
+          newWarpOp.getLaneid(), newWarpOp.getWarpSize(), innerWarpInputVals,
+          innerWarpInputTypes);
+
+      innerWarp.getWarpRegion().takeBody(*oldIfBranch->getParent());
+      innerWarp.getWarpRegion().addArguments(
+          innerWarpInputTypes,
+          SmallVector<Location>(innerWarpInputTypes.size(), ifOp.getLoc()));
+
+      SmallVector<Value> yieldOperands;
+      for (Value operand : oldIfBranch->getTerminator()->getOperands())
+        yieldOperands.push_back(operand);
+      rewriter.eraseOp(oldIfBranch->getTerminator());
+
+      rewriter.setInsertionPointToEnd(innerWarp.getBody());
+      gpu::YieldOp::create(rewriter, innerWarp.getLoc(), yieldOperands);
+      rewriter.setInsertionPointAfter(innerWarp);
+      scf::YieldOp::create(rewriter, ifOp.getLoc(), innerWarp.getResults());
+
+      // Update any users of escaping values that were forwarded to the
+      // inner `WarpOp`. These values are now arguments of the inner `WarpOp`.
+      innerWarp.walk([&](Operation *op) {
+        for (OpOperand &operand : op->getOpOperands()) {
+          auto it = escapeValToBlockArgIndex.find(operand.get());
+          if (it == escapeValToBlockArgIndex.end())
+            continue;
+          operand.set(innerWarp.getBodyRegion().getArgument(it->second));
+        }
+      });
+      mlir::vector::moveScalarUniformCode(innerWarp);
+    };
+    processBranch(&ifOp.getThenRegion().front(),
+                  &newIfOp.getThenRegion().front(), escapingValuesThen,
+                  escapingValueInputTypesThen);
+    if (!ifOp.getElseRegion().empty())
+      processBranch(&ifOp.getElseRegion().front(),
+                    &newIfOp.getElseRegion().front(), escapingValuesElse,
+                    escapingValueInputTypesElse);
+    // Update the users of `<- WarpOp.yield <- IfOp.yield` to use the new `IfOp`
+    // result.
+    for (auto [origIdx, newIdx] : ifResultMapping)
+      rewriter.replaceAllUsesExcept(warpOp.getResult(origIdx),
+                                    newIfOp.getResult(newIdx), newIfOp);
+    // Similarly, update any users of the `WarpOp` results that were not
+    // results of the `IfOp`.
+    for (auto [origIdx, newIdx] : origToNewYieldIdx)
+      rewriter.replaceAllUsesWith(warpOp.getResult(origIdx),
+                                  newWarpOp.getResult(newIdx));
+    // Remove the original `WarpOp` and `IfOp`, they should not have any uses
+    // at this point.
+    rewriter.eraseOp(ifOp);
+    rewriter.eraseOp(warpOp);
+    return success();
+  }
+
+private:
+  DistributionMapFn distributionMapFn;
+};
+
 /// Sink scf.for region out of WarpExecuteOnLane0Op. This can be done only if
 /// the scf.ForOp is the last operation in the region so that it doesn't
 /// change the order of execution. This creates a new scf.for region after the
@@ -2068,6 +2267,8 @@ void mlir::vector::populatePropagateWarpVectorDistributionPatterns(
                                     benefit);
   patterns.add<WarpOpScfForOp>(patterns.getContext(), distributionMapFn,
                                benefit);
+  patterns.add<WarpOpScfIfOp>(patterns.getContext(), distributionMapFn,
+                              benefit);
 }
 
 void mlir::vector::populateDistributeReduction(
diff --git a/mlir/test/Dialect/Vector/vector-warp-distribute.mlir b/mlir/test/Dialect/Vector/vector-warp-distribute.mlir
index 8750582ef1e1f..bb7639204022f 100644
--- a/mlir/test/Dialect/Vector/vector-warp-distribute.mlir
+++ b/mlir/test/Dialect/Vector/vector-warp-distribute.mlir
@@ -1856,3 +1856,72 @@ func.func @negative_warp_step_more_than_warp_size(%laneid: index, %buffer: memre
 // CHECK-PROP-LABEL: @negative_warp_step_more_than_warp_size
 // CHECK-PROP-NOT: vector.broadcast
 // CHECK-PROP: vector.step : vector<64xindex>
+
+// -----
+
+func.func @warp_scf_if_no_yield_distribute(%buffer: memref<128xindex>, %pred : i1)  {
+  %laneid = gpu.lane_id
+  %c0 = arith.constant 0 : index
+
+  gpu.warp_execute_on_lane_0(%laneid)[32] {
+    %seq = vector.step : vector<32xindex>
+    scf.if %pred {
+      vector.store %seq, %buffer[%c0] : memref<128xindex>, vector<32xindex>
+    }
+    gpu.yield
+  }
+  return
+}
+
+// CHECK-PROP-LABEL: func.func @warp_scf_if_no_yield_distribute(
+//  CHECK-PROP-SAME:   %[[ARG0:.+]]: memref<128xindex>, %[[ARG1:.+]]: i1
+//       CHECK-PROP:   scf.if %[[ARG1]] {
+//       CHECK-PROP:   gpu.warp_execute_on_lane_0(%{{.*}})[32] args(%{{.*}} : vector<1xindex>) {
+//       CHECK-PROP:   ^bb0(%[[ARG2:.+]]: vector<32xindex>):
+//       CHECK-PROP:   vector.store %[[ARG2]], %[[ARG0]][%{{.*}}] : memref<128xindex>, vector<32xindex>
+
+// -----
+
+func.func @warp_scf_if_distribute(%pred : i1)  {
+  %laneid = gpu.lane_id
+  %c0 = arith.constant 0 : index
+
+  %0 = gpu.warp_execute_on_lane_0(%laneid)[32] -> vector<1xf32> {
+    %seq1 = vector.step : vector<32xindex>
+    %seq2 = arith.constant dense<2> : vector<32xindex>
+    %0 = scf.if %pred -> (vector<32xf32>) {
+      %1 = "some_op"(%seq1) : (vector<32xindex>) -> (vector<32xf32>)
+      scf.yield %1 : vector<32xf32>
+    } else {
+      %2 = "other_op"(%seq2) : (vector<32xindex>) -> (vector<32xf32>)
+      scf.yield %2 : vector<32xf32>
+    }
+    gpu.yield %0 : vector<32xf32>
+  }
+  "some_use"(%0) : (vector<1xf32>) -> ()
+
+  return
+}
+
+// CHECK-PROP-LABEL: func.func @warp_scf_if_distribute(
+//  CHECK-PROP-SAME:    %[[ARG0:.+]]: i1
+//       CHECK-PROP:    %[[SEQ2:.+]] = arith.constant dense<2> : vector<32xindex>
+//       CHECK-PROP:    %[[LANE_ID:.+]] = gpu.lane_id
+//       CHECK-PROP:    %[[SEQ1:.+]] = vector.broadcast %[[LANE_ID]] : index to vector<1xindex>
+//       CHECK-PROP:    %[[IF_YIELD_DIST:.+]] = scf.if %[[ARG0]] -> (vector<1xf32>) {
+//       CHECK-PROP:    %[[THEN_DIST:.+]] = gpu.warp_execute_on_lane_0(%[[LANE_ID]])[32] args(%[[SEQ1]] : vector<1xindex>) -> (vector<1xf32>) {
+//       CHECK-PROP:        ^bb0(%[[ARG1:.+]]: vector<32xindex>):
+//       CHECK-PROP:        %{{.*}} = "some_op"(%[[ARG1]]) : (vector<32xindex>) -> vector<32xf32>
+//       CHECK-PROP:        gpu.yield %{{.*}} : vector<32xf32>
+//       CHECK-PROP:      }
+//       CHECK-PROP:      scf.yield %[[THEN_DIST]] : vector<1xf32>
+//       CHECK-PROP:    } else {
+//       CHECK-PROP:      %[[ELSE_DIST:.+]] = gpu.warp_execute_on_lane_0(%[[LANE_ID]])[32] -> (vector<1xf32>) {
+//       CHECK-PROP:        %{{.*}} = "other_op"(%[[SEQ2]]) : (vector<32xindex>) -> vector<32xf32>
+//       CHECK-PROP:        gpu.yield %{{.*}} : vector<32xf32>
+//       CHECK-PROP:      }
+//       CHECK-PROP:      scf.yield %[[ELSE_DIST]] : vector<1xf32>
+//       CHECK-PROP:    }
+//       CHECK-PROP:    "some_use"(%[[IF_YIELD_DIST]]) : (vector<1xf32>) -> ()
+//       CHECK-PROP:    return
+//       CHECK-PROP:  }

[Garra1980]

cc @charithaintc

[charithaintc]

overall this looks great to me. awesome work.

I added some comments. I will do another pass after your response.

[charithaintc]

LGTM. please address the remainnig comments before merge.