[QualcommQnn] support models (#9412)

lite/backends/nnadapter/nnadapter/include/nnadapter/operation/math/expand.h

@@ -0,0 +1,51 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <vector>
+#include "operation/math/utility.h"
+
+namespace nnadapter {
+namespace operation {
+namespace math {
+
+template <typename T>
+static int expand(const T* input_data,
+                  const std::vector<int32_t>& input_shape,
+                  T* output_data,
+                  const std::vector<int32_t>& output_shape) {
+  std::vector<int> in_stride(input_shape.size(), 1);
+  std::vector<int> out_stride(output_shape.size(), 1);
+  for (int i = input_shape.size() - 2; i >= 0; --i) {
+    in_stride[i] = input_shape[i + 1] * in_stride[i + 1];
+  }
+  for (int i = output_shape.size() - 2; i >= 0; --i) {
+    out_stride[i] = output_shape[i + 1] * out_stride[i + 1];
+  }
+  int out_size = shape_production(output_shape);
+  for (int out_id = 0; out_id < out_size; ++out_id) {
+    int in_id = 0;
+    for (int i = input_shape.size() - 1; i >= 0; --i) {
+      int in_j = (out_id / out_stride[i]) % input_shape[i];
+      in_id += in_j * in_stride[i];
+    }
+    output_data[out_id] = input_data[in_id];
+  }
+  return 0;
+}
+
+}  // namespace math
+}  // namespace operation
+}  // namespace nnadapter

lite/backends/nnadapter/nnadapter/include/nnadapter/optimizer/convert_datalayout_nchw_to_nhwc.h

@@ -39,6 +39,7 @@ class NCHW2NHWCDataLayoutConverter {
   void ConvertAdaptivePool2D(core::Operation* operation);
   void ConvertBatchNormalization(core::Operation* operation);
   void ConvertCast(core::Operation* operation);
+  void ConvertChannelShuffle(core::Operation* operation);
   void ConvertClip(core::Operation* operation);
   void ConvertComparisons(core::Operation* operation);
   void ConvertCumSum(core::Operation* operation);
@@ -55,6 +56,7 @@ class NCHW2NHWCDataLayoutConverter {
   void ConvertLeakyRelu(core::Operation* operation);
   void ConvertLpNormalization(core::Operation* operation);
   void ConvertActivation(core::Operation* operation);
+  void ConvertPad(core::Operation* operation);
   void ConvertPow(core::Operation* operation);
   void ConvertQuantize(core::Operation* operation);
   void ConvertReduce(core::Operation* operation);

lite/backends/nnadapter/nnadapter/include/nnadapter/optimizer/fuse_unsqueeze_pad_squeeze_into_pad.h

@@ -0,0 +1,23 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "core/types.h"
+
+namespace nnadapter {
+
+void FuseUnsqueezePadSqueezeIntoPad(core::Model *model);
+
+}  // namespace nnadapter

lite/backends/nnadapter/nnadapter/include/nnadapter/optimizer/pattern_matcher.h

@@ -98,6 +98,8 @@ class PatternMatcher {
                                      int index = -1);
     Pattern* IsOperationOutputOperand(NNAdapterOperationType type,
                                       int index = -1);
+    Pattern* HasInLinksNum(int num);
+    Pattern* HasOutLinksNum(int num);
     // Mark the pattern matched node to be deleted, so its inlinks and outlinks
     // should be inside a matched subgraph.
     Pattern* IsIntermediate();

lite/backends/nnadapter/nnadapter/src/driver/huawei_ascend_npu/converter/all.h

@@ -26,8 +26,8 @@ REGISTER_CONVERTER(ASSIGN, ConvertAssign)
 REGISTER_CONVERTER(AVERAGE_POOL_2D, ConvertPool2D)
 REGISTER_CONVERTER(BATCH_NORMALIZATION, ConvertBatchNormalization)
 REGISTER_CONVERTER(CAST, ConvertCast)
-REGISTER_CONVERTER(CLIP, ConvertClip)
 REGISTER_CONVERTER(CHANNEL_SHUFFLE, ConvertChannelShuffle)
+REGISTER_CONVERTER(CLIP, ConvertClip)
 REGISTER_CONVERTER(CONCAT, ConvertConcat)
 REGISTER_CONVERTER(CONV_2D, ConvertConv2D)
 REGISTER_CONVERTER(CONV_2D_TRANSPOSE, ConvertConv2DTranspose)

lite/backends/nnadapter/nnadapter/src/operation/expand.cc

@@ -15,6 +15,7 @@
 #include "operation/expand.h"
 #include <vector>
 #include "core/types.h"
+#include "operation/math/expand.h"
 #include "utility/debug.h"
 #include "utility/hints.h"
 #include "utility/logging.h"
@@ -70,7 +71,7 @@ NNADAPTER_EXPORT void UpdateExpandInferOutputShape(
 }
 
 NNADAPTER_EXPORT bool ValidateExpand(const core::Operation* operation) {
-  return false;
+  return true;
 }
 
 NNADAPTER_EXPORT int PrepareExpand(core::Operation* operation) {
@@ -120,7 +121,50 @@ NNADAPTER_EXPORT int PrepareExpand(core::Operation* operation) {
 }
 
 NNADAPTER_EXPORT int ExecuteExpand(core::Operation* operation) {
-  return NNADAPTER_FEATURE_NOT_SUPPORTED;
+  EXPAND_OPERATION_EXTRACT_INPUTS_OUTPUTS
+
+  auto in_dims_data = input_operand->type.dimensions.data;
+  auto in_dims_count = input_operand->type.dimensions.count;
+  std::vector<int32_t> in_dims(in_dims_data, in_dims_data + in_dims_count);
+  auto out_dims_data = input_operand->type.dimensions.data;
+  auto out_dims_count = input_operand->type.dimensions.count;
+  std::vector<int32_t> out_dims(out_dims_data, out_dims_data + out_dims_count);
+  auto in_dtype_length =
+      GetOperandPrecisionDataLength(input_operand->type.precision);
+  auto output_buffer = AllocateOperand(output_operand);
+  int status = -1;
+  switch (in_dtype_length) {
+    case 1:
+      status = math::expand(reinterpret_cast<int8_t*>(input_operand->buffer),
+                            in_dims,
+                            static_cast<int8_t*>(output_buffer),
+                            out_dims);
+      break;
+    case 2:
+      status = math::expand(reinterpret_cast<int16_t*>(input_operand->buffer),
+                            in_dims,
+                            static_cast<int16_t*>(output_buffer),
+                            out_dims);
+      break;
+    case 4:
+      status = math::expand(reinterpret_cast<int32_t*>(input_operand->buffer),
+                            in_dims,
+                            static_cast<int32_t*>(output_buffer),
+                            out_dims);
+      break;
+    case 8:
+      status = math::expand(reinterpret_cast<int64_t*>(input_operand->buffer),
+                            in_dims,
+                            static_cast<int64_t*>(output_buffer),
+                            out_dims);
+      break;
+    default:
+      NNADAPTER_LOG(FATAL) << "Not support data type length: "
+                           << in_dtype_length;
+      break;
+  }
+  NNADAPTER_CHECK_EQ(status, 0);
+  return NNADAPTER_NO_ERROR;
 }
 
 }  // namespace operation

lite/backends/nnadapter/nnadapter/src/optimizer/convert_datalayout_nchw_to_nhwc.cc

@@ -249,6 +249,22 @@ void NCHW2NHWCDataLayoutConverter::ConvertCast(core::Operation* operation) {
   SetPermutation(output_operand, input_permutation);
 }
 
+void NCHW2NHWCDataLayoutConverter::ConvertChannelShuffle(
+    core::Operation* operation) {
+  auto& input_operands = operation->input_operands;
+  auto& output_operands = operation->output_operands;
+  auto input_count = input_operands.size();
+  auto output_count = output_operands.size();
+  NNADAPTER_CHECK_EQ(input_count, 2);
+  NNADAPTER_CHECK_EQ(output_count, 1);
+  auto input_operand = input_operands[0];
+  auto output_operand = output_operands[0];
+  // The input and output operands share the same dimorder vector
+  auto input_permutation = GetPermutation(input_operand);
+  TransposeOperand(output_operand, input_permutation);
+  SetPermutation(output_operand, input_permutation);
+}
+
 void NCHW2NHWCDataLayoutConverter::ConvertClip(core::Operation* operation) {
   auto& input_operands = operation->input_operands;
   auto& output_operands = operation->output_operands;
@@ -632,6 +648,31 @@ void NCHW2NHWCDataLayoutConverter::ConvertLpNormalization(
   SetPermutation(output_operand, input_permutation);
 }
 
+void NCHW2NHWCDataLayoutConverter::ConvertPad(core::Operation* operation) {
+  auto& input_operands = operation->input_operands;
+  auto& output_operands = operation->output_operands;
+  auto input_count = input_operands.size();
+  auto output_count = output_operands.size();
+  NNADAPTER_CHECK_EQ(input_count, 4);
+  NNADAPTER_CHECK_EQ(output_count, 1);
+  auto input_operand = input_operands[0];
+  auto output_operand = output_operands[0];
+  // The input and output operands share the same dimorder vector
+  auto input_permutation = GetPermutation(input_operand);
+  TransposeOperand(output_operand, input_permutation);
+  SetPermutation(output_operand, input_permutation);
+  // Trans pads
+  auto pads_data = reinterpret_cast<int32_t*>(input_operands[1]->buffer);
+  int32_t pads_size = input_operands[1]->length / sizeof(int32_t);
+  NNADAPTER_CHECK_EQ(input_permutation.size() * 2, pads_size);
+  std::vector<int32_t> trans_pads_data(pads_size);
+  for (int i = 0; i < input_permutation.size(); i++) {
+    trans_pads_data[i * 2] = pads_data[input_permutation[i] * 2];
+    trans_pads_data[i * 2 + 1] = pads_data[input_permutation[i] * 2 + 1];
+  }
+  memcpy(pads_data, trans_pads_data.data(), pads_size * sizeof(int32_t));
+}
+
 void NCHW2NHWCDataLayoutConverter::ConvertPow(core::Operation* operation) {
   auto& input_operands = operation->input_operands;
   auto& output_operands = operation->output_operands;
@@ -1218,6 +1259,9 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
       case NNADAPTER_CAST:
         ConvertCast(operation);
         break;
+      case NNADAPTER_CHANNEL_SHUFFLE:
+        ConvertChannelShuffle(operation);
+        break;
       case NNADAPTER_CLIP:
         ConvertClip(operation);
         break;
@@ -1271,6 +1315,9 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
       case NNADAPTER_MAT_MUL:
         ConvertMatMul(operation);
         break;
+      case NNADAPTER_PAD:
+        ConvertPad(operation);
+        break;
       case NNADAPTER_POW:
         ConvertPow(operation);
         break;

lite/backends/nnadapter/nnadapter/src/optimizer/convert_quantization_symm_to_asymm.cc

@@ -115,6 +115,7 @@ NNADAPTER_EXPORT void ConvertQuantizationSymmToAsymm(core::Model* model) {
       case NNADAPTER_LAYER_NORMALIZATION:
       case NNADAPTER_LEAKY_RELU:
       case NNADAPTER_MAX_POOL_2D:
+      case NNADAPTER_PAD:
       case NNADAPTER_RELU:
       case NNADAPTER_RELU6:
       case NNADAPTER_RESHAPE:

lite/backends/nnadapter/nnadapter/src/optimizer/fuse_reshape_transpose_reshape_into_channel_shuffle.cc

@@ -12,10 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "optimizer/fuse_reshape_transpose_reshape_into_channel_shuffle.h"
 #include <algorithm>
 #include <map>
 #include <vector>
-#include "optimizer/fuse_conv2d_activation_into_conv2d.h"
 #include "optimizer/pattern_matcher.h"
 #include "utility/debug.h"
 #include "utility/logging.h"
@@ -91,8 +91,7 @@ void ReshapeTransposeReshapeFuser::BuildPattern() {
           ->IsIntermediate();
   auto last_reshape_output_pattern =
       CreatePattern("last_reshape_output")
-          ->IsOperationOutputOperand(NNADAPTER_RESHAPE, 0)
-          ->IsIntermediate();
+          ->IsOperationOutputOperand(NNADAPTER_RESHAPE, 0);
   // Create the topological connections for the above patterns
   std::vector<Pattern*> first_reshape_input_patterns{
       first_reshape_input_pattern, first_reshape_shape_pattern};