microsoft · titaiwangms · Jan 16, 2025 · Jan 11, 2025 · Jan 11, 2025 · Jan 13, 2025
diff --git a/docs/How_To_Update_ONNX_Dev_Notes.md b/docs/How_To_Update_ONNX_Dev_Notes.md
@@ -36,10 +36,10 @@ This file should be generated. See [cgmanifests/README](/cgmanifests/README.md)
 1. If there is a build failure in stage "Check out of dated documents" in WebAssembly CI pipeline, update ONNX Runtime
    Web WebGL operator support document:
    - Make sure Node.js is installed (see [Prerequisites](../js/README.md#Prerequisites) for instructions).
-   - Follow step 1 in [js/Build](../js/README.md#Build-2) to install dependencies).
+   - Follow [js/Build](../js/README.md#Build-2) to install dependencies.
    - Follow instructions in [Generate document](../js/README.md#Generating-Document) to update document. Commit changes applied to file `docs/operators.md`.
 
-1. Usually some newly introduced tests will fail. Then you may need to update
+2. Usually some newly introduced tests will fail. Then you may need to update
 - [onnxruntime/test/onnx/main.cc](/onnxruntime/test/onnx/main.cc)
 - [onnxruntime/test/providers/cpu/model_tests.cc](/onnxruntime/test/providers/cpu/model_tests.cc)
 - [csharp/test/Microsoft.ML.OnnxRuntime.Tests.NetCoreApp/InferenceTest.netcore.cs](/csharp/test/Microsoft.ML.OnnxRuntime.Tests.NetCoreApp/InferenceTest.netcore.cs)

diff --git a/docs/OperatorKernels.md b/docs/OperatorKernels.md
diff --git a/onnxruntime/core/providers/cpu/activation/activations.cc b/onnxruntime/core/providers/cpu/activation/activations.cc
@@ -33,8 +33,10 @@ namespace onnxruntime {
       op, since_version, type,                                           \
       KernelDefBuilder().MayInplace(0, 0).TypeConstraint("T", DataTypeImpl::GetTensorType<type>()), op<type>);
 
-REGISTER_UNARY_ELEMENTWISE_KERNEL(Elu, 6);
-REGISTER_UNARY_ELEMENTWISE_KERNEL(HardSigmoid, 6);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(Elu, 6, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(Elu, 22);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(HardSigmoid, 6, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(HardSigmoid, 22);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(LeakyRelu, 6, 15);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Relu, 6, 12, float);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Relu, 6, 12, double);
@@ -52,19 +54,23 @@ REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(LeakyRelu, 6, 15, MLFloat16);
 REGISTER_UNARY_ELEMENTWISE_TYPED_KERNEL(LeakyRelu, 16, MLFloat16);
 #endif  // MLAS_F16VEC_INTRINSICS_SUPPORTED
 
-REGISTER_UNARY_ELEMENTWISE_KERNEL(Selu, 6);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(Selu, 6, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(Selu, 22);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Sigmoid, 6, 12, float);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Sigmoid, 6, 12, double);
 REGISTER_UNARY_ELEMENTWISE_TYPED_KERNEL(Sigmoid, 13, float);
 REGISTER_UNARY_ELEMENTWISE_TYPED_KERNEL(Sigmoid, 13, double);
-REGISTER_UNARY_ELEMENTWISE_KERNEL(Softplus, 1);
-REGISTER_UNARY_ELEMENTWISE_KERNEL(Softsign, 1);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(Softplus, 1, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(Softplus, 22);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(Softsign, 1, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(Softsign, 22);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Tanh, 6, 12, float);
 REGISTER_VERSIONED_UNARY_ELEMENTWISE_TYPED_KERNEL(Tanh, 6, 12, double);
 REGISTER_UNARY_ELEMENTWISE_KERNEL(Celu, 12);
 REGISTER_UNARY_ELEMENTWISE_TYPED_KERNEL(Tanh, 13, float);
 REGISTER_UNARY_ELEMENTWISE_TYPED_KERNEL(Tanh, 13, double);
-REGISTER_UNARY_ELEMENTWISE_KERNEL(ThresholdedRelu, 10);
+REGISTER_VERSIONED_UNARY_ELEMENTWISE_KERNEL(ThresholdedRelu, 10, 21);
+REGISTER_UNARY_ELEMENTWISE_KERNEL(ThresholdedRelu, 22);
 
 // Opset-16 adds BFloat16 to allowed types for the LeakyRelu operator
 REGISTER_UNARY_ELEMENTWISE_KERNEL(LeakyRelu, 16);

diff --git a/onnxruntime/core/providers/cpu/cpu_execution_provider.cc b/onnxruntime/core/providers/cpu/cpu_execution_provider.cc
diff --git a/onnxruntime/core/providers/cpu/fp16/fp16_conv.cc b/onnxruntime/core/providers/cpu/fp16/fp16_conv.cc
@@ -596,9 +596,15 @@ Status FusedConvFp16::Compute(OpKernelContext* context) const {
 // Operator definitions
 //
 
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    Conv,
+    11, 21, MLFloat16,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
+    FusedConvFp16);
+
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     Conv,
-    11,
+    22,
     MLFloat16,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     FusedConvFp16);

diff --git a/onnxruntime/core/providers/cpu/fp16/fp16_pool.cc b/onnxruntime/core/providers/cpu/fp16/fp16_pool.cc
@@ -224,9 +224,16 @@ ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     PoolFp16);
 
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    MaxPool,
+    12, 21,
+    MLFloat16,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
+    PoolFp16);
+
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     MaxPool,
-    12,
+    22,
     MLFloat16,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     PoolFp16);
@@ -237,16 +244,30 @@ ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     PoolFp16);
 
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    AveragePool,
+    19, 21,
+    MLFloat16,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
+    PoolFp16);
+
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     AveragePool,
-    19,
+    22,
+    MLFloat16,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
+    PoolFp16);
+
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    GlobalAveragePool,
+    1, 21,
     MLFloat16,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     PoolFp16);
 
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     GlobalAveragePool,
-    1,
+    22,
     MLFloat16,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()),
     PoolFp16);

diff --git a/onnxruntime/core/providers/cpu/math/det.cc b/onnxruntime/core/providers/cpu/math/det.cc
@@ -13,9 +13,17 @@ using namespace onnxruntime::common;
 
 namespace onnxruntime {
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Det,
     11,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Det<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Det,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Det<float>);
 

diff --git a/onnxruntime/core/providers/cpu/math/element_wise_ops.cc b/onnxruntime/core/providers/cpu/math/element_wise_ops.cc
@@ -1337,16 +1337,30 @@ class Sin final : public OpKernel {
   }
 };
 
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    Sin,
+    7, 21,
+    float,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Sin<float>);
+
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(
+    Sin,
+    7, 21,
+    double,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<double>()),
+    Sin<double>);
+
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     Sin,
-    7,
+    22,
     float,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Sin<float>);
 
 ONNX_CPU_OPERATOR_TYPED_KERNEL(
     Sin,
-    7,
+    22,
     double,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<double>()),
     Sin<double>);
@@ -1365,9 +1379,17 @@ class Cos final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Cos,
     7,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Cos<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Cos,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Cos<float>);
 
@@ -1385,9 +1407,15 @@ class Tan final : public OpKernel {
   }
 };
 
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
+    Tan,
+    7, 21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Tan<float>);
+
 ONNX_CPU_OPERATOR_KERNEL(
     Tan,
-    7,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Tan<float>);
 
@@ -1405,9 +1433,17 @@ class Asin final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Asin,
     7,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Asin<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Asin,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Asin<float>);
 
@@ -1425,9 +1461,17 @@ class Acos final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Acos,
     7,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Acos<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Acos,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Acos<float>);
 
@@ -1445,9 +1489,17 @@ class Atan final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Atan,
     7,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Atan<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Atan,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Atan<float>);
 
@@ -1465,9 +1517,17 @@ class Sinh final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Sinh,
     9,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Sinh<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Sinh,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Sinh<float>);
 
@@ -1485,9 +1545,17 @@ class Cosh final : public OpKernel {
   }
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Cosh,
     9,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Cosh<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Cosh,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Cosh<float>);
 
@@ -1517,9 +1585,17 @@ class Asinh final : public OpKernel {
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Asinh);
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Asinh,
     9,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Asinh<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Asinh,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Asinh<float>);
 
@@ -1549,9 +1625,17 @@ class Acosh final : public OpKernel {
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Acosh);
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Acosh,
     9,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Acosh<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Acosh,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Acosh<float>);
 
@@ -1581,9 +1665,17 @@ class Atanh final : public OpKernel {
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Atanh);
 };
 
-ONNX_CPU_OPERATOR_KERNEL(
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
     Atanh,
     9,
+    21,
+    KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
+    Atanh<float>);
+
+// Opset 22 starts to support bfloat16
+ONNX_CPU_OPERATOR_KERNEL(
+    Atanh,
+    22,
     KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()),
     Atanh<float>);
 

diff --git a/onnxruntime/core/providers/cpu/math/round.cc b/onnxruntime/core/providers/cpu/math/round.cc
@@ -12,9 +12,13 @@
 
 namespace onnxruntime {
 
-ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 11, MLFloat16, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()), Round<MLFloat16>);
-ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 11, float, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()), Round<float>);
-ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 11, double, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<double>()), Round<double>);
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(Round, 11, 21, MLFloat16, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()), Round<MLFloat16>);
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(Round, 11, 21, float, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()), Round<float>);
+ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(Round, 11, 21, double, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<double>()), Round<double>);
+
+ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 22, MLFloat16, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<MLFloat16>()), Round<MLFloat16>);
+ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 22, float, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<float>()), Round<float>);
+ONNX_CPU_OPERATOR_TYPED_KERNEL(Round, 22, double, KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<double>()), Round<double>);
 
 template <typename T>
 Status Round<T>::Compute(OpKernelContext* ctx) const {

diff --git a/onnxruntime/core/providers/cpu/nn/Unpool.cc b/onnxruntime/core/providers/cpu/nn/Unpool.cc
@@ -23,9 +23,17 @@ ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
         .TypeConstraint("T2", DataTypeImpl::GetTensorType<int64_t>()),
     MaxUnpool);
 
+ONNX_CPU_OPERATOR_VERSIONED_KERNEL(
+    MaxUnpool,
+    11, 21,
+    KernelDefBuilder()
+        .TypeConstraint("T1", DataTypeImpl::GetTensorType<float>())
+        .TypeConstraint("T2", DataTypeImpl::GetTensorType<int64_t>()),
+    MaxUnpool);
+
 ONNX_CPU_OPERATOR_KERNEL(
     MaxUnpool,
-    11,
+    22,
     KernelDefBuilder()
         .TypeConstraint("T1", DataTypeImpl::GetTensorType<float>())
         .TypeConstraint("T2", DataTypeImpl::GetTensorType<int64_t>()),