[QNN-EP] Support for Upsample operator

zhaoxul-qti · zhaoxul-qti · commit e7d7194835bd · 2025-04-01T19:04:56.000+08:00
diff --git a/onnxruntime/core/providers/qnn/builder/op_builder_factory.cc b/onnxruntime/core/providers/qnn/builder/op_builder_factory.cc
@@ -132,6 +132,10 @@ OpBuilderRegistrations::OpBuilderRegistrations() {
     CreateResizeOpBuilder("Resize", *this);
   }
 
+  {
+    CreateUpsampleOpBuilder("Upsample", *this);
+  }
+
   {
     CreateTopKOpBuilder("TopK", *this);
   }
diff --git a/onnxruntime/core/providers/qnn/builder/op_builder_factory.h b/onnxruntime/core/providers/qnn/builder/op_builder_factory.h
@@ -75,6 +75,8 @@ void CreateSplitOpBuilder(const std::string& op_type, OpBuilderRegistrations& op
 
 void CreateResizeOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
 
+void CreateUpsampleOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
+
 void CreateTopKOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
 
 void CreateTileOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations);
diff --git a/onnxruntime/core/providers/qnn/builder/opbuilder/base_op_builder.h b/onnxruntime/core/providers/qnn/builder/opbuilder/base_op_builder.h
@@ -187,6 +187,7 @@ class BaseOpBuilder : public IOpBuilder {
 
         {"Reshape", QNN_OP_RESHAPE},
         {"Resize", QNN_OP_RESIZE},
+        {"Upsample", QNN_OP_RESIZE},
         {"Flatten", QNN_OP_RESHAPE},
         {"Squeeze", QNN_OP_RESHAPE},
         {"Unsqueeze", QNN_OP_RESHAPE},
diff --git a/onnxruntime/core/providers/qnn/builder/opbuilder/upsample_op_builder.cc b/onnxruntime/core/providers/qnn/builder/opbuilder/upsample_op_builder.cc
@@ -0,0 +1,243 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include <cassert>
+#include <unordered_map>
+
+#include "core/providers/qnn/builder/opbuilder/base_op_builder.h"
+#include "core/providers/qnn/builder/qnn_utils.h"
+#include "core/providers/qnn/builder/qnn_model_wrapper.h"
+#include "core/providers/qnn/builder/op_builder_factory.h"
+
+namespace onnxruntime {
+namespace qnn {
+
+class UpsampleOpBuilder : public BaseOpBuilder {
+ public:
+  UpsampleOpBuilder() : BaseOpBuilder("UpsampleOpBuilder") {}
+  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(UpsampleOpBuilder);
+
+  Status IsOpSupported(QnnModelWrapper& qnn_model_wrapper,
+                       const NodeUnit& node_unit,
+                       const logging::Logger& logger) const override final ORT_MUST_USE_RESULT;
+
+ protected:
+  Status ProcessInputs(QnnModelWrapper& qnn_model_wrapper,
+                       const NodeUnit& node_unit,
+                       const logging::Logger& logger,
+                       std::vector<std::string>& input_names,
+                       bool do_op_validation) const override ORT_MUST_USE_RESULT;
+
+  Status ProcessAttributesAndOutputs(QnnModelWrapper& qnn_model_wrapper,
+                                     const NodeUnit& node_unit,
+                                     std::vector<std::string>&& input_names,
+                                     const logging::Logger& logger,
+                                     bool do_op_validation) const override ORT_MUST_USE_RESULT;
+
+  Status OverrideOutputQuantParam(QnnModelWrapper& qnn_model_wrapper,
+                                  const NodeUnit& node_unit,
+                                  const logging::Logger& logger,
+                                  const std::vector<std::string>& input_names,
+                                  size_t output_index,
+                                  Qnn_DataType_t qnn_data_type,
+                                  QnnQuantParamsWrapper& quant_param) const override ORT_MUST_USE_RESULT;
+
+ private:
+  const std::unordered_map<std::string, uint32_t> supported_modes = {
+    {"nearest", QNN_OP_RESIZE_INTERPOLATION_MODE_NEAREST},
+    {"linear", QNN_OP_RESIZE_INTERPOLATION_MODE_LINEAR},
+    {"cubic", QNN_OP_RESIZE_INTERPOLATION_MODE_CUBIC}};
+
+  // Info for Onnx Upsample attribute {"<attribute_name>", <default_value>}
+  const OnnxAttrInfo<std::string> onnx_mode_attr = {"mode", "nearest"};
+};
+
+static Status AddQnnScalar(QnnModelWrapper& qnn_model_wrapper,
+                           const NodeUnit& node_unit,
+                           std::vector<std::string>& param_tensor_names,
+                           const Qnn_Scalar_t& qnn_scalar,
+                           const std::string& qnn_scalar_param_name) {
+  QnnParamWrapper qnn_param_wrapper(node_unit.Index(), node_unit.Name(), qnn_scalar_param_name, qnn_scalar);
+  param_tensor_names.push_back(qnn_param_wrapper.GetParamTensorName());
+  qnn_model_wrapper.AddParamWrapper(std::move(qnn_param_wrapper));
+
+  return Status::OK();
+}
+
+Status UpsampleOpBuilder::IsOpSupported(QnnModelWrapper& qnn_model_wrapper,
+                                        const NodeUnit& node_unit,
+                                        const logging::Logger& logger) const {
+  // Resize ops are sensitive with data layout, no special validation so far
+  // The nodes from 1st call of GetCapability do not get layout transformer applied, it's still NCHW
+  // The nodes from 2nd call of GetCapability get layout transformer applied, it's NHWC
+  // Need to do op validation in 1st call of GetCapability
+  if (node_unit.Domain() == kMSInternalNHWCDomain) {
+    return AddToModelBuilder(qnn_model_wrapper, node_unit, logger, true);
+  }
+
+  const bool is_npu_backend = IsNpuBackend(qnn_model_wrapper.GetQnnBackendType());
+  NodeAttrHelper node_helper(node_unit);
+
+  // Check mode
+  const std::string interp_mode = GetOnnxAttr(node_helper, onnx_mode_attr);
+  ORT_RETURN_IF_NOT(supported_modes.find(interp_mode) != supported_modes.end(),
+                    "QNN EP: Resize does not support mode ", interp_mode.c_str());
+
+  const auto& input_0 = node_unit.Inputs()[0];
+  std::vector<uint32_t> input_shape;
+  ORT_RETURN_IF_NOT(qnn_model_wrapper.GetOnnxShape(input_0.node_arg, input_shape),
+                    "QNN EP: Cannot get input shape for Onnx Upsample ", input_0.node_arg.Name().c_str());
+  const size_t input_rank = input_shape.size();
+
+  ORT_RETURN_IF(is_npu_backend && (input_rank < 3 || input_rank > 5),
+                "QNN EP: The input rank for Resize must be at least 3 and no greater than 5 on the HTP.");
+
+  const auto& output_0 = node_unit.Outputs()[0];
+  std::vector<uint32_t> output_shape;
+  ORT_RETURN_IF_NOT(qnn_model_wrapper.GetOnnxShape(output_0.node_arg, output_shape),
+                    "QNN EP: Cannot get output shape for Onnx Upsample ", output_0.node_arg.Name().c_str(),
+                    ". Dynamic scales input is not supported in QNN EP.");
+
+  // Check that only the spatial dimensions (width, height) are resized. The batch_size (N) and channels (C) should
+  // be untouched. This code runs before layout transformation, so we know that the current layout is "channel first"
+  // (e.g., N, C, S1, S2, ..., SN).
+  ORT_RETURN_IF_NOT(input_shape[0] == output_shape[0] && input_shape[1] == output_shape[1],
+                    "QNN EP: Resize may only change the spatial dimensions.");
+
+  return Status::OK();
+}
+
+Status UpsampleOpBuilder::ProcessInputs(QnnModelWrapper& qnn_model_wrapper,
+                                        const NodeUnit& node_unit,
+                                        const logging::Logger& logger,
+                                        std::vector<std::string>& input_names,
+                                        bool do_op_validation) const {
+  ORT_UNUSED_PARAMETER(do_op_validation);
+
+  const auto& inputs = node_unit.Inputs();
+
+  // Only need to consider the first input of Onnx upsample.
+  ORT_RETURN_IF_ERROR(ProcessInput(qnn_model_wrapper, inputs[0], logger, input_names));
+
+  return Status::OK();
+}
+
+Status UpsampleOpBuilder::ProcessAttributesAndOutputs(QnnModelWrapper& qnn_model_wrapper,
+                                                      const NodeUnit& node_unit,
+                                                      std::vector<std::string>&& input_names,
+                                                      const logging::Logger& logger,
+                                                      bool do_op_validation) const {
+  std::vector<std::string> param_tensor_names;
+  NodeAttrHelper node_helper(node_unit);
+  const std::string interp_mode = GetOnnxAttr(node_helper, onnx_mode_attr);
+
+  const auto& input_0 = node_unit.Inputs()[0];
+  std::vector<uint32_t> input_shape;
+  ORT_RETURN_IF_NOT(qnn_model_wrapper.GetOnnxShape(input_0.node_arg, input_shape),
+                    "QNN EP: Cannot get input shape for Onnx Upsample ", input_0.node_arg.Name().c_str());
+
+  const size_t input_rank = input_shape.size();
+  const bool is_npu_backend = IsNpuBackend(qnn_model_wrapper.GetQnnBackendType());
+  std::string qnn_op_type = QNN_OP_RESIZE;
+
+  if (is_npu_backend && input_rank == 4 && interp_mode != "cubic") {
+    // Translate QNN's Resize to QNN's ResizeNearestNeighbor/ResizeBilinear to achieve better performance on
+    // the HTP backend. QNN's ResizeNearestNeighbor and ResizeBilinear are only supported when input rank is 4.
+    qnn_op_type = (interp_mode == "nearest") ? QNN_OP_RESIZE_NEAREST_NEIGHBOR : QNN_OP_RESIZE_BILINEAR;
+
+    // Parameter 'align_corners'
+    Qnn_Scalar_t qnn_align_corners = QNN_SCALAR_INIT;
+    qnn_align_corners.dataType = QNN_DATATYPE_BOOL_8;
+    qnn_align_corners.bool8Value = false;
+    const std::string align_corners_param_name = (qnn_op_type == QNN_OP_RESIZE_BILINEAR)
+                                                 ? QNN_OP_RESIZE_BILINEAR_PARAM_ALIGN_CORNERS
+                                                 : QNN_OP_RESIZE_NEAREST_NEIGHBOR_PARAM_ALIGN_CORNERS;
+
+    ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                     qnn_align_corners, align_corners_param_name));
+
+    // Parameter 'half_pixel_centers'
+    Qnn_Scalar_t qnn_half_pixel_centers = QNN_SCALAR_INIT;
+    qnn_half_pixel_centers.dataType = QNN_DATATYPE_BOOL_8;
+    qnn_half_pixel_centers.bool8Value = false;
+    const std::string half_pixel_centers_param_name = (qnn_op_type == QNN_OP_RESIZE_BILINEAR)
+                                                      ? QNN_OP_RESIZE_BILINEAR_PARAM_HALF_PIXEL_CENTERS
+                                                      : QNN_OP_RESIZE_NEAREST_NEIGHBOR_PARAM_HALF_PIXEL_CENTERS;
+
+    ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                     qnn_half_pixel_centers, half_pixel_centers_param_name));
+
+    if (qnn_op_type == QNN_OP_RESIZE_BILINEAR) {
+      // Parameter 'antialias'
+      Qnn_Scalar_t qnn_antialias = QNN_SCALAR_INIT;
+      qnn_antialias.dataType = QNN_DATATYPE_BOOL_8;
+      qnn_antialias.bool8Value = false;
+
+      ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                       qnn_antialias, QNN_OP_RESIZE_BILINEAR_PARAM_ANTIALIAS));
+    }
+  }
+  else {
+    // Remain as QNN's Resize
+    // Parameter 'exclude_outside'
+    Qnn_Scalar_t qnn_exclude_outside = QNN_SCALAR_INIT;
+    qnn_exclude_outside.dataType = QNN_DATATYPE_BOOL_8;
+    qnn_exclude_outside.bool8Value = false;
+
+    ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                     qnn_exclude_outside, QNN_OP_RESIZE_PARAM_EXCLUDE_OUTSIDE));
+
+    // Parameter 'transformation_mode'
+    Qnn_Scalar_t qnn_transformation_mode = QNN_SCALAR_INIT;
+    qnn_transformation_mode.dataType = QNN_DATATYPE_UINT_32;
+    qnn_transformation_mode.uint32Value = QNN_OP_RESIZE_TRANSFORMATION_MODE_ASYMMETRIC;
+
+    ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                     qnn_transformation_mode, QNN_OP_RESIZE_PARAM_TRANSFORMATION_MODE));
+
+    // Parameter 'interpolation_mode'
+    Qnn_Scalar_t qnn_interp_mode = QNN_SCALAR_INIT;
+    qnn_interp_mode.dataType = QNN_DATATYPE_UINT_32;
+    qnn_interp_mode.uint32Value = supported_modes.at(interp_mode);
+
+    ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                     qnn_interp_mode, QNN_OP_RESIZE_PARAM_INTERPOLATION_MODE));
+
+    // Parameter 'nearest_mode'. Process only when 'interpolation_mode' is NEAREST.
+    if (qnn_interp_mode.uint32Value == QNN_OP_RESIZE_INTERPOLATION_MODE_NEAREST) {
+      Qnn_Scalar_t qnn_nearest_mode = QNN_SCALAR_INIT;
+      qnn_nearest_mode.dataType = QNN_DATATYPE_UINT_32;
+      qnn_nearest_mode.uint32Value = QNN_OP_RESIZE_NEAREST_MODE_ROUND_PREFER_FLOOR;
+
+      ORT_RETURN_IF_ERROR(AddQnnScalar(qnn_model_wrapper, node_unit, param_tensor_names,
+                                       qnn_nearest_mode, QNN_OP_RESIZE_PARAM_NEAREST_MODE));
+    }
+  }
+
+  return ProcessOutputs(qnn_model_wrapper, node_unit, std::move(input_names), std::move(param_tensor_names),
+                        logger, do_op_validation, qnn_op_type);
+}
+
+Status UpsampleOpBuilder::OverrideOutputQuantParam(QnnModelWrapper& qnn_model_wrapper,
+                                                   const NodeUnit& node_unit,
+                                                   const logging::Logger& logger,
+                                                   const std::vector<std::string>& input_names,
+                                                   size_t output_index,
+                                                   Qnn_DataType_t qnn_data_type,
+                                                   QnnQuantParamsWrapper& quant_param) const {
+  if (!quant_param.IsPerTensor()) {
+    return Status::OK();
+  }
+
+  // Force Resize op's output to use the same quantization parameters as the input if nearly equal.
+  // This helps the HTP backend employ certain optimizations.
+  return SetOutputQParamEqualToInputIfNearlyEqual(qnn_model_wrapper, node_unit, logger, input_names,
+                                                  0 /*input_index*/, output_index, qnn_data_type, quant_param);
+}
+
+void CreateUpsampleOpBuilder(const std::string& op_type, OpBuilderRegistrations& op_registrations) {
+  op_registrations.AddOpBuilder(op_type, std::make_unique<UpsampleOpBuilder>());
+}
+
+}  // namespace qnn
+}  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/qnn/upsample_op_test.cc b/onnxruntime/test/providers/qnn/upsample_op_test.cc

Original file line number	Diff line number	Diff line change
`@@ -132,6 +132,10 @@ OpBuilderRegistrations::OpBuilderRegistrations() {`
`132`	`132`	`CreateResizeOpBuilder("Resize", *this);`
`133`	`133`	`}`
`134`	`134`
	`135`	`+ {`
	`136`	`+ CreateUpsampleOpBuilder("Upsample", *this);`
	`137`	`+ }`
	`138`	`+`
`135`	`139`	`{`
`136`	`140`	`CreateTopKOpBuilder("TopK", *this);`
`137`	`141`	`}`