From b667cb1f6c3207dcfa54d478edc71e4abc491c6e Mon Sep 17 00:00:00 2001
From: Wanming Lin <wanming.lin@intel.com>
Date: Thu, 11 Jun 2026 16:09:53 +0800
Subject: [PATCH] [WebNN EP] Remove unnecessary cast around normalization ops
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove fp16→fp32→fp16 cast workaround from SimplifiedLayerNormalization
and SkipSimplifiedLayerNormalization. The WebNN backend now handles
float16 precision correctly, making these intermediate casts unnecessary.
---
 .../builders/impl/normalization_op_builder.cc | 44 +------------------
 1 file changed, 2 insertions(+), 42 deletions(-)

diff --git a/onnxruntime/core/providers/webnn/builders/impl/normalization_op_builder.cc b/onnxruntime/core/providers/webnn/builders/impl/normalization_op_builder.cc
index c9cf7bb162870..2d1dd3d15cc7b 100644
--- a/onnxruntime/core/providers/webnn/builders/impl/normalization_op_builder.cc
+++ b/onnxruntime/core/providers/webnn/builders/impl/normalization_op_builder.cc
@@ -96,33 +96,9 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
       ORT_RETURN_IF_NOT(GetType(*input_defs[0], input_type, logger), "Cannot get input type");
       emscripten::val common_options = emscripten::val::object();
 
-      if (input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
-        // Decomposed *SimplifiedLayerNormalization may lose precision if its data type is float16.
-        // So cast all inputs to float32 to ensure precision.
-        common_options.set("label", node.Name() + "_cast_input_to_fp32");
-        input = model_builder.GetBuilder().call<emscripten::val>("cast", input,
-                                                                 emscripten::val("float32"), common_options);
-
-        common_options.set("label", node.Name() + "_cast_scale_to_fp32");
-        scale = model_builder.GetBuilder().call<emscripten::val>("cast", scale,
-                                                                 emscripten::val("float32"), common_options);
-
-        if (!bias.isUndefined()) {
-          common_options.set("label", node.Name() + "_cast_bias_to_fp32");
-          bias = model_builder.GetBuilder().call<emscripten::val>("cast", bias,
-                                                                  emscripten::val("float32"), common_options);
-        }
-      }
-
       // If it is SkipSimplifiedLayerNormalization, add the skip and bias (if it exists) to the input.
       if (op_type == "SkipSimplifiedLayerNormalization") {
         emscripten::val skip = model_builder.GetOperand(input_defs[1]->Name());
-        if (input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
-          // Cast skip to float32
-          common_options.set("label", node.Name() + "_cast_skip_to_fp32");
-          skip = model_builder.GetBuilder().call<emscripten::val>("cast", skip,
-                                                                  emscripten::val("float32"), common_options);
-        }
         common_options.set("label", node.Name() + "_add_skip");
         input = model_builder.GetBuilder().call<emscripten::val>("add", input, skip, common_options);
         if (!bias.isUndefined()) {
@@ -134,20 +110,12 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
         // Now input equals to input_skip_bias_sum.
         if (TensorExists(output_defs, 3)) {
           emscripten::val input_skip_bias_sum = input;
-          if (input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
-            // Cast input_skip_bias_sum back to float16.
-            common_options.set("label", node.Name() + "_cast_input_skip_bias_sum_to_fp16");
-            input_skip_bias_sum = model_builder.GetBuilder().call<emscripten::val>("cast", input_skip_bias_sum,
-                                                                                   emscripten::val("float16"),
-                                                                                   common_options);
-          }
           model_builder.AddOperand(output_defs[3]->Name(), input_skip_bias_sum);
         }
       }
 
       // Pow
-      emscripten::val pow_constant =
-          model_builder.CreateOrGetConstant<float>(ONNX_NAMESPACE::TensorProto_DataType_FLOAT, 2);
+      emscripten::val pow_constant = model_builder.CreateOrGetConstant<float>(input_type, 2);
       common_options.set("label", node.Name() + "_pow");
       emscripten::val pow =
           model_builder.GetBuilder().call<emscripten::val>("pow", input, pow_constant, common_options);
@@ -160,8 +128,7 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
       emscripten::val reduce_mean = model_builder.GetBuilder().call<emscripten::val>("reduceMean", pow, reduce_options);
 
       // Add
-      emscripten::val add_constant =
-          model_builder.CreateOrGetConstant<float>(ONNX_NAMESPACE::TensorProto_DataType_FLOAT, epsilon);
+      emscripten::val add_constant = model_builder.CreateOrGetConstant<float>(input_type, epsilon);
       common_options.set("label", node.Name() + "_add");
       emscripten::val add =
           model_builder.GetBuilder().call<emscripten::val>("add", reduce_mean, add_constant, common_options);
@@ -183,13 +150,6 @@ Status NormalizationOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder
         common_options.set("label", node.Name() + "_add_bias");
         output = model_builder.GetBuilder().call<emscripten::val>("add", output, bias, common_options);
       }
-
-      if (input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
-        // Cast output back to float16.
-        common_options.set("label", node.Name() + "_cast_output_to_fp16");
-        output = model_builder.GetBuilder().call<emscripten::val>("cast", output,
-                                                                  emscripten::val("float16"), common_options);
-      }
     }
   } else if (op_type == "InstanceNormalization") {
     // WebNN spec only supports 4D input for instanceNormalization.