[js/webgpu]FusedConv

js/web/lib/wasm/jsep/webgpu/op-resolve-rules.ts

@@ -50,6 +50,7 @@ export const WEBGPU_OP_RESOLVE_RULES: Map<string, OperatorImplementation> = new
   ['Clip', [unaryOps.clip]],
   ['Concat', [concat, parseConcatAttributes]],
   ['Conv', [conv, parseConvAttributes]],
+  ['FusedConv', [conv, parseConvAttributes]],
   ['ConvTranspose', [convTranspose, parseConvTransposeAttributes]],
   ['Cos', [unaryOps.cos]],
   ['Cosh', [unaryOps.cosh]],

js/web/test/data/ops/fused-conv.jsonc

@@ -0,0 +1,67 @@
+[
+  {
+    "name": "Fused Conv Relu",
+    "operator": "FusedConv",
+    "attributes": [
+      { "name": "kernel_shape", "data": [2, 2], "type": "ints" },
+      { "name": "strides", "data": [1, 1], "type": "ints" },
+      { "name": "pads", "data": [0, 0, 0, 0], "type": "ints" },
+      { "name": "group", "data": 1, "type": "int" },
+      { "name": "dilations", "data": [1, 1], "type": "ints" },
+      { "name": "activation", "data": "Relu", "type": "string" }
+    ],
+    "cases": [
+      {
+        "name": "T[0]",
+        "inputs": [
+          {
+            "data": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
+            "dims": [1, 1, 3, 3],
+            "type": "float32"
+          },
+          {
+            "data": [1.0, 1.0, 1.0, 1.0, -1.0, -1.0, -1.0, -1.0],
+            "dims": [2, 1, 2, 2],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [12.0, 16.0, 24.0, 28.0, 0.0, 0.0, 0.0, 0.0],
+            "dims": [1, 2, 2, 2],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
+  },
+  {
+    "name": "conv without bias addition A",
+    "operator": "Conv",
+    "attributes": [{ "name": "kernel_shape", "data": [2, 2], "type": "ints" }],
+    "cases": [
+      {
+        "name": "T[0]",
+        "inputs": [
+          {
+            "data": [10, 20, 30, 40, 50, 60, 70, 80, 90],
+            "dims": [1, 1, 3, 3],
+            "type": "float32"
+          },
+          {
+            "data": [1, 2, 3, 4],
+            "dims": [1, 1, 2, 2],
+            "type": "float32"
+          }
+        ],
+        "outputs": [
+          {
+            "data": [370, 470, 670, 770],
+            "dims": [1, 1, 2, 2],
+            "type": "float32"
+          }
+        ]
+      }
+    ]
+  }
+]

onnxruntime/core/providers/js/js_execution_provider.cc

@@ -240,6 +240,8 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSInternalNHW
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSInternalNHWCDomain, 1, float, GlobalAveragePool);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSInternalNHWCDomain, 1, float, GlobalMaxPool);
 
+class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSDomain, 1, float, FusedConv);
+
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 1, 10, float, Conv);
 class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 11, float, Conv);
 class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 1, 10, float, ConvTranspose);
@@ -500,6 +502,8 @@ std::unique_ptr<KernelRegistry> RegisterKernels() {
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSInternalNHWCDomain, 1, float, GlobalAveragePool)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kMSInternalNHWCDomain, 1, float, GlobalMaxPool)>,
 
+      BuildKernelCreateInfo<ONNX_CPU_OPERATOR_TYPED_MS_KERNEL(kJsExecutionProvider, kMSDomain, 1, float, FusedConv)>,
+
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 1, 10, float, Conv)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 11, float, Conv)>,
       BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kJsExecutionProvider, kOnnxDomain, 1, 10, float, ConvTranspose)>,

onnxruntime/core/providers/js/operators/conv.cc

@@ -33,7 +33,15 @@ namespace js {
       T,                                                                                   \
       kJsExecutionProvider,                                                                \
       (*KernelDefBuilder::Create()).TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
-      Conv<T, false>);
+      Conv<T, false>);                                                                     \
+  ONNX_OPERATOR_TYPED_KERNEL_EX(                                                           \
+      FusedConv,                                                                           \
+      kMSDomain,                                                                           \
+      1,                                                                                   \
+      T,                                                                                   \
+      kJsExecutionProvider,                                                                \
+      (*KernelDefBuilder::Create()).TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
+      FusedConv<T, false>);
 
 REGISTER_KERNEL_TYPED(float)
 

onnxruntime/core/providers/js/operators/conv.h

@@ -94,5 +94,12 @@ class Conv : public JsKernel {
   // Tensor w_transposed_;
 };
 
+template <typename T, bool is_channels_last>
+class FusedConv : public Conv<T, is_channels_last> {
+ public:
+  explicit FusedConv(const OpKernelInfo& info) : Conv<T, is_channels_last>(info) {
+    ORT_ENFORCE(info.GetAttr<std::string>("activation", &(this->conv_attrs_.activation)).IsOK());
+  }
+};
 }  // namespace js
 }  // namespace onnxruntime