Merge branch 'main' into mschofie/windows-fixes

Author: Sunny-Anand

src/Dialect/ONNX/Transforms/Recompose.cpp

@@ -271,8 +271,6 @@ struct RecomposeLayerNormFromMulPattern : public OpRewritePattern<ONNXMulOp> {
       return reportFailure("RMS norm mul has too many uses");
     if (isdRecipOp && !isdRecipOp->hasOneUse())
       return reportFailure("RMS norm recip has too many uses");
-    if (!nsMulOp->hasOneUse())
-      return reportFailure("RMS norm scale mul has too many uses");
     // Now check values epsilon.
     if (!isScalarTensor(epsilon))
       return reportFailure("RMS epsilon is expected to be scalar");

test/mlir/onnx/onnx_recompose.mlir

@@ -247,6 +247,28 @@ func.func @rms_layer_norm_v2(%x: tensor<1x384x768xf32>, %scale: tensor<768xf32>,
 
 // -----
 
+// RMS Layer norm with multiple uses of the scale multiplication
+func.func @rms_layer_norm_multi_use(%x: tensor<1x384x768xf32>, %scale: tensor<768xf32>) -> (tensor<1x384x768xf32>) {
+  %eps = onnx.Constant dense<9.99999974E-6> : tensor<f32>
+  %dd = "onnx.Mul"(%x, %x) : (tensor<1x384x768xf32>, tensor<1x384x768xf32>) -> tensor<1x384x768xf32>
+  %var = "onnx.ReduceMeanV13"(%dd) {axes = [-1], keepdims = 1 : si64, onnx_node_name = "ReduceMean_42"} : (tensor<1x384x768xf32>) -> tensor<1x384x1xf32>
+  %varEps = "onnx.Add"(%eps, %var) : (tensor<f32>, tensor<1x384x1xf32>) -> tensor<1x384x1xf32>
+  %StdDev = "onnx.Sqrt"(%varEps) : (tensor<1x384x1xf32>) -> tensor<1x384x1xf32>
+  %Norm = "onnx.Div"(%x, %StdDev) : (tensor<1x384x768xf32>, tensor<1x384x1xf32>) -> tensor<1x384x768xf32>
+  %NormScaled = "onnx.Mul"(%scale, %Norm) : (tensor<768xf32>, tensor<1x384x768xf32>) -> tensor<1x384x768xf32>
+  %MultiUse = "onnx.Add"(%NormScaled, %NormScaled) : (tensor<1x384x768xf32>, tensor<1x384x768xf32>) -> tensor<1x384x768xf32>
+  return %MultiUse : tensor<1x384x768xf32>
+// CHECK-LABEL:  func.func @rms_layer_norm_multi_use
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1x384x768xf32>, [[PARAM_1_:%.+]]: tensor<768xf32>) -> tensor<1x384x768xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.NoValue"() {value} : () -> none
+// CHECK:           [[VAR_Y_:%.+]], [[VAR_InvStdDev_:%.+]] = "onnx.RMSLayerNormalization"([[PARAM_0_]], [[PARAM_1_]], [[VAR_0_]]) {axis = 2 : si64, epsilon = 9.99999974E-6 : f32, stash_type = 1 : si64} : (tensor<1x384x768xf32>, tensor<768xf32>, none) -> (tensor<1x384x768xf32>, none)
+// CHECK:           [[VAR_1_:%.+]] = "onnx.Add"([[VAR_Y_]], [[VAR_Y_]]) : (tensor<1x384x768xf32>, tensor<1x384x768xf32>) -> tensor<1x384x768xf32>
+// CHECK:           return [[VAR_1_]] : tensor<1x384x768xf32>
+// CHECK:         }
+}
+
+// -----
+
 // COM: QLinearMatMul
 func.func @qlinear_matmul(%arg0: tensor<?x?x768xi8>, %arg1: tensor<f32>, %arg2: tensor<i8>, %arg3: tensor<768x768xi8>, %arg4: tensor<f32>, %arg5: tensor<i8>, %arg6: tensor<f32>, %arg7: tensor<i8>) -> (tensor<?x?x768xi8>) {
     %0 = "onnx.DequantizeLinear"(%arg0, %arg1, %arg2) {axis = 1 : si64} : (tensor<?x?x768xi8>, tensor<f32>, tensor<i8>) -> tensor<?x?x768xf32>