[WebNN] Accept Float16Array for float16 data type if it is available (#23894)

Float16Array is now shipping and WebNN Chromium implementation has
accepted it. We should allow it in WebNN EP as well.

Author: Honry

js/web/lib/wasm/jsep/backend-webnn.ts

@@ -314,7 +314,8 @@ export class WebNNBackend {
         bufferView = new Float32Array(buffer);
         break;
       case 'float16':
-        bufferView = new Uint16Array(buffer);
+        bufferView =
+          typeof Float16Array !== 'undefined' && Float16Array.from ? new Float16Array(buffer) : new Uint16Array(buffer);
         break;
       case 'int32':
         bufferView = new Int32Array(buffer);

onnxruntime/core/providers/webnn/builders/impl/rotaryEmbedding_op_builder.cc

@@ -219,9 +219,17 @@ Status RotaryEmbeddingOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_build
     sign_buffer.set(0, -1.0f);
     sign_buffer.set(1, 1.0f);
   } else if (input_data_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
-    sign_buffer = emscripten::val::global("Uint16Array").new_(2);
-    sign_buffer.set(0, PackFloat32ToUint16AsFloat16(-1.0f));
-    sign_buffer.set(1, PackFloat32ToUint16AsFloat16(1.0f));
+    if (model_builder.IsFloat16ArrayAvailable()) {
+      // Float16Array is avaliable - use Float16Array.
+      sign_buffer = emscripten::val::global("Float16Array").new_(2);
+      sign_buffer.set(0, -1.0f);
+      sign_buffer.set(1, 1.0f);
+    } else {
+      // Float16Array is not available - use Uint16Array instead.
+      sign_buffer = emscripten::val::global("Uint16Array").new_(2);
+      sign_buffer.set(0, PackFloat32ToUint16AsFloat16(-1.0f));
+      sign_buffer.set(1, PackFloat32ToUint16AsFloat16(1.0f));
+    }
   } else {
     return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "Unsupported input data type: ", input_data_type);
   }

onnxruntime/core/providers/webnn/builders/model_builder.cc

@@ -197,7 +197,8 @@ Status ModelBuilder::RegisterInitializers() {
 
         // Wasm memory grow will cause all array buffers reallocation, which will be treated as detached
         // buffers in JS side. Simply create a copy to fix it.
-        operand = wnn_builder_.call<emscripten::val>("constant", desc, view.call<emscripten::val>("slice"));
+        view = view.call<emscripten::val>("slice");
+        operand = wnn_builder_.call<emscripten::val>("constant", desc, view["buffer"]);
       }
     } else {
       // TODO: support other type.
@@ -350,7 +351,8 @@ Status ModelBuilder::AddOperandFromPersistMemoryBuffer(
   emscripten::val operand = emscripten::val::object();
   // Wasm memory grow will cause all array buffers reallocation, which will be treated as detached
   // buffers in JS side. Simply create a copy to fix it.
-  operand = wnn_builder_.call<emscripten::val>("constant", desc, view.call<emscripten::val>("slice"));
+  view = view.call<emscripten::val>("slice");
+  operand = wnn_builder_.call<emscripten::val>("constant", desc, view["buffer"]);
 
   AddOperand(name, operand);
   mem_persist_buffers_.push_back(std::move(persist_buffer));

onnxruntime/core/providers/webnn/builders/model_builder.h

@@ -30,6 +30,7 @@ class ModelBuilder {
   Status Compile(std::unique_ptr<Model>& model) ORT_MUST_USE_RESULT;
 
   // Accessors for members.
+  bool IsFloat16ArrayAvailable() const { return is_float16array_available_; }
   const GraphViewer& GetGraphViewer() const { return graph_viewer_; }
   InitializedTensorSet GetInitializerTensors();
 
@@ -68,6 +69,8 @@ class ModelBuilder {
  private:
   const GraphViewer& graph_viewer_;
   const logging::Logger& logger_;
+  const bool is_float16array_available_ = !emscripten::val::global("Float16Array").isUndefined() &&
+                                          emscripten::val::global("Float16Array").hasOwnProperty("from");
 
   emscripten::val wnn_context_ = emscripten::val::undefined();
   emscripten::val wnn_builder_ = emscripten::val::undefined();
@@ -172,9 +175,12 @@ const emscripten::val& ModelBuilder::CreateOrGetConstant(const int32_t& data_typ
         }
         break;
       case ONNX_NAMESPACE::TensorProto_DataType_FLOAT16:
-        buffer = emscripten::val::global("Uint16Array").new_(num_elements);
+        buffer = is_float16array_available_
+                     ? emscripten::val::global("Float16Array").new_(num_elements)
+                     : emscripten::val::global("Uint16Array").new_(num_elements);
         if (value) {
-          buffer.call<void>("fill", emscripten::val(PackFloat32ToUint16AsFloat16(value)));
+          buffer.call<void>("fill",
+                            emscripten::val(is_float16array_available_ ? value : PackFloat32ToUint16AsFloat16(value)));
         }
         break;
       case ONNX_NAMESPACE::TensorProto_DataType_FLOAT: