[js/webgpu] enable f16 for concat (#18528)

### Description With this PR `realesrgan-t64-f16` models becomes 32.8 ms from 1052.55 ms. Now the whole model run on jsep.
microsoft · Nov 21, 2023 · ac8598a · ac8598a
1 parent 81a763a
commit ac8598a
Showing 1 changed file with 8 additions and 4 deletions.
diff --git a/onnxruntime/core/providers/js/operators/concat.cc b/onnxruntime/core/providers/js/operators/concat.cc
@@ -12,7 +12,8 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
     1, 3,
     kJsExecutionProvider,
     (*KernelDefBuilder::Create())
-        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
+        .TypeConstraint("T", {DataTypeImpl::GetTensorType<MLFloat16>(),
+                              DataTypeImpl::GetTensorType<float>(),
                               DataTypeImpl::GetTensorType<int32_t>()}),
     Concat);
 
@@ -22,7 +23,8 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
     4, 10,
     kJsExecutionProvider,
     (*KernelDefBuilder::Create())
-        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
+        .TypeConstraint("T", {DataTypeImpl::GetTensorType<MLFloat16>(),
+                              DataTypeImpl::GetTensorType<float>(),
                               DataTypeImpl::GetTensorType<int32_t>()}),
     Concat);
 
@@ -32,7 +34,8 @@ ONNX_OPERATOR_VERSIONED_KERNEL_EX(
     11, 12,
     kJsExecutionProvider,
     (*KernelDefBuilder::Create())
-        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
+        .TypeConstraint("T", {DataTypeImpl::GetTensorType<MLFloat16>(),
+                              DataTypeImpl::GetTensorType<float>(),
                               DataTypeImpl::GetTensorType<int32_t>()}),
     Concat);
 
@@ -42,7 +45,8 @@ ONNX_OPERATOR_KERNEL_EX(
     13,
     kJsExecutionProvider,
     (*KernelDefBuilder::Create())
-        .TypeConstraint("T", {DataTypeImpl::GetTensorType<float>(),
+        .TypeConstraint("T", {DataTypeImpl::GetTensorType<MLFloat16>(),
+                              DataTypeImpl::GetTensorType<float>(),
                               DataTypeImpl::GetTensorType<int32_t>()}),
     Concat);