simplifier scalar tensor shape to scalar i32 (#1310)

* simplfier scalar tensor shape to scalar i32

* fix ut

* update

Author: Yancey0623

tao_compiler/mlir/disc/transforms/disc_algebraic_simplifier.cc

@@ -9,6 +9,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <stack>
+
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/Support/Debug.h"
@@ -423,7 +425,6 @@ struct BroadCastInDimOfReshapeOpCanonicalizationPattern
     return success();
   }
 };
-
 // Simplifier extract and from-element op pattern, an example as following:
 //  %0 = tensor.extract %arg0[] : tensor<f32>
 //  %1 = tensor.from_elements %0 : tensor<1xf32>
@@ -439,6 +440,8 @@ struct SimplifierFromElementsPattern
     auto loc = op->getLoc();
     Value input = op->getOperand(0);
     Value result = op->getResult(0);
+    // only support scalar tensor
+    if (op->getNumOperands() != 1) return failure();
     auto extractOp = input.getDefiningOp<tensor::ExtractOp>();
     if (!extractOp) return failure();
 
@@ -530,6 +533,93 @@ struct IndexCastSimplifierPattern
     return failure();
   }
 };
+// Simplify get_dimension_size pattern. An examples as following:
+// Case 1):
+// %2 = "mhlo.get_dimension_size"(%1)
+// %3 = "tensor.extract" %2[] -> i32
+// %from_elements = tensor.from_elements %3, ...
+// Convert to:
+// %2 = "tensor.dim"(%1, %cst0) -> i32
+// %from_elements = tensor.from_elements %2, ...
+//
+// Case 2):
+// %2 = "mhlo.get_dimension_size"(%1)
+// %3 = mhlo.mul %2, %4
+// %4 = "tensor.extract" %3[] -> i32
+// %from_elements = tensor.from_elements %4, ...
+// Convert to:
+// %2 = "tensor.dim"(%1, %cst0) -> i32
+// %3 = arith.mul %2, %4
+// %from_elements = tensor.from_elements %3, ...
+
+struct SimplifierGetDimensionSizePattern
+    : public OpRewritePattern<mhlo::GetDimensionSizeOp> {
+  using OpRewritePattern<mhlo::GetDimensionSizeOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(mhlo::GetDimensionSizeOp getDimOp,
+                                PatternRewriter& rewriter) const override {
+    auto loc = getDimOp->getLoc();
+    Value tensor = getDimOp->getOperand(0);
+    auto dim = getDimOp.getDimension();
+    auto elemTy = getDimOp.getResult()
+                      .getType()
+                      .cast<RankedTensorType>()
+                      .getElementType();
+
+    SmallVector<Operation*, 4> ops;
+    std::stack<Operation*> stack;
+    for (auto user : getDimOp->getUsers()) {
+      if (isa<tensor::ExtractOp, func::ReturnOp, mhlo::ReshapeOp,
+              mhlo::DynamicBroadcastInDimOp>(user))
+        continue;
+      stack.push(user);
+    }
+    while (!stack.empty()) {
+      auto user = stack.top();
+      stack.pop();
+      ops.push_back(user);
+      for (auto op : user->getUsers()) {
+        if (isa<tensor::ExtractOp, func::ReturnOp, mhlo::ReshapeOp,
+                mhlo::DynamicBroadcastInDimOp>(op))
+          continue;
+        stack.push(op);
+      }
+    }
+    for (auto op : ops) {
+      auto loc = op->getLoc();
+      rewriter.setInsertionPoint(op);
+      auto v1 = rewriter.create<tensor::ExtractOp>(loc, op->getOperand(0));
+      auto v2 = rewriter.create<tensor::ExtractOp>(loc, op->getOperand(1));
+      Value newOpValue;
+      if (isa<mhlo::MulOp>(op)) {
+        newOpValue = rewriter.create<arith::MulIOp>(loc, v1, v2).getResult();
+      } else if (isa<mhlo::AddOp>(op)) {
+        newOpValue = rewriter.create<arith::AddIOp>(loc, v1, v2).getResult();
+      } else if (isa<mhlo::SubtractOp>(op)) {
+        newOpValue = rewriter.create<arith::SubIOp>(loc, v1, v2).getResult();
+      } else if (isa<mhlo::DivOp>(op)) {
+        newOpValue = rewriter.create<arith::DivSIOp>(loc, v1, v2).getResult();
+      } else {
+        return failure();
+      }
+      auto result = rewriter.create<tensor::FromElementsOp>(
+          loc, getDimOp.getResult().getType().cast<RankedTensorType>(),
+          newOpValue);
+      op->replaceAllUsesWith(result);
+    }
+    rewriter.setInsertionPoint(getDimOp);
+    auto dimValue =
+        rewriter.create<tensor::DimOp>(loc, tensor, dim).getResult();
+    auto castValue = rewriter.create<arith::IndexCastOp>(loc, elemTy, dimValue);
+    auto dimValueTensor =
+        rewriter
+            .create<tensor::FromElementsOp>(
+                loc, getDimOp.getResult().getType().cast<RankedTensorType>(),
+                ValueRange{castValue})
+            .getResult();
+    getDimOp.replaceAllUsesWith(dimValueTensor);
+    return success();
+  }
+};
 
 // Consant folding the broadcasted constant, for patterns like:
 //   %0 = mhlo.constant // Scalar or splat constant
@@ -627,7 +717,8 @@ void populateDiscAlgebraicSimplifierPatterns(RewritePatternSet& patterns) {
     IdentityBroadCastInDimOpCanonicalizationPattern<mhlo::DynamicBroadcastInDimOp>,
     SimplifierFromElementsPattern,
     TrunciSimplifierPattern,
-    IndexCastSimplifierPattern
+    IndexCastSimplifierPattern,
+    SimplifierGetDimensionSizePattern
   >(patterns.getContext());
   if (isMemIntensiveOptExperimentalEnabled()) {
     // Will be enabled by default after a set of robustness testing.

tao_compiler/mlir/disc/transforms/tests/disc-algebraic-simplifier.mlir

@@ -256,4 +256,21 @@ func.func @select_simp(%arg0: tensor<16xf16>) -> (tensor<20xf16>, tensor<20xf16>
   %10 = mhlo.constant dense<false> : tensor<20xi1>
   %11 = "mhlo.select"(%10, %7, %3): (tensor<20xi1>, tensor<20xf16>, tensor<20xf16>) -> tensor<20xf16>
   return %8, %3 : tensor<20xf16>, tensor<20xf16>
-}
+}
+
+// -----
+
+// CHECK-LABEL: @main
+func.func @main(%arg0: tensor<?x10xf32>, %arg1: tensor<10xf32>) -> tensor<?x10xf32> {
+  %c10_i32 = arith.constant 10 : i32
+  %c_0 = mhlo.constant dense<4> : tensor<i32>
+  // CHECK: %dim = tensor.dim %arg0, %c0 : tensor<?x10xf32>
+  // CHECK: %0 = arith.index_cast %dim : index to i32
+  %2 = "mhlo.get_dimension_size"(%arg0) {dimension = 0 : i64} : (tensor<?x10xf32>) -> tensor<i32>
+  // CHECK: %1 = arith.muli %0, %c4_i32 : i32
+  %3 = mhlo.multiply %2, %c_0 : tensor<i32>
+  %extracted = tensor.extract %3[] : tensor<i32>
+  %from_elements = tensor.from_elements %extracted, %c10_i32 : tensor<2xi32>
+  %4 = "mhlo.dynamic_broadcast_in_dim"(%arg1, %from_elements) {broadcast_dimensions = dense<1> : tensor<1xi64>} : (tensor<10xf32>, tensor<2xi32>) -> tensor<?x10xf32>
+  return %4 : tensor<?x10xf32>
+}