[QualcommQnn] add tile; cast, mul, tile pass

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

@@ -20,4 +20,6 @@ namespace nnadapter {
 
 void ConstantFoldShapeAndAssociatedOperations(core::Model* model);
 
+void FoldShapeSliceConcatTile(core::Model* model);
+
 }  // namespace nnadapter

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

@@ -67,6 +67,7 @@ class NCHW2NHWCDataLayoutConverter {
   void ConvertSplit(core::Operation* operation);
   void ConvertSqueeze(core::Operation* operation);
   void ConvertStack(core::Operation* operation);
+  void ConvertTile(core::Operation* operation);
   void ConvertTranspose(core::Operation* operation);
   void ConvertMatMul(core::Operation* operation);
   void ConvertUnsqueeze(core::Operation* operation);

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

@@ -87,7 +87,7 @@ class PatternMatcher {
     bool MatchAllConditions(const Node* node) const;
     // Utility conditions
     Pattern* IsOperand();
-    Pattern* IsOperation(NNAdapterOperationType type);
+    Pattern* IsOperation(NNAdapterOperationType type = NNADAPTER_UNKNOWN);
     Pattern* IsConstantOperand();
     Pattern* IsVariableOperand();
     Pattern* IsConstantCopyOperand();

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

@@ -0,0 +1,25 @@
+// 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 RemoveUselessCast(core::Model *model);
+
+void RemoveUselessMul(core::Model *model);
+
+}  // namespace nnadapter

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

@@ -193,11 +193,8 @@ void FillLikeCumSumSubGatherAddFuser::BuildPattern() {
       CreatePattern("cum_sum_axis")
           ->IsOperationInputOperand(NNADAPTER_CUM_SUM, 1)
           ->MatchCondition([](const Node* node) -> bool {
-            auto operand = node->operand;
-            return operand != nullptr &&
-                   operand->type.precision == NNADAPTER_INT32 &&
-                   operand->length == sizeof(int32_t) &&
-                   *reinterpret_cast<int32_t*>(operand->buffer) == 1;
+            int32_t axis = *reinterpret_cast<int32_t*>(node->operand->buffer);
+            return axis == 1 || axis == -1;
           })
           ->IsIntermediate();
   auto cum_sum_exclusive =

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

@@ -14,6 +14,7 @@
 
 #include "optimizer/constant_fold_shape_and_associated_operations.h"
 #include <set>
+#include "optimizer/pattern_matcher.h"
 #include "utility/debug.h"
 #include "utility/hints.h"
 #include "utility/logging.h"
@@ -132,4 +133,111 @@ NNADAPTER_EXPORT void ConstantFoldShapeAndAssociatedOperations(
   }
 }
 
+/*
+before:
+  shape    fill_like    fill_like
+    |          |            |
+  slice        |            |
+    \----------|------------/
+            concat
+               |
+             tile
+
+after:
+   tile
+*/
+class FoldShapeSliceConcatTileFuser : public PatternMatcher {
+ public:
+  FoldShapeSliceConcatTileFuser() {}
+  void BuildPattern() override;
+  bool HandleMatchedResults(
+      core::Model *model, const std::map<std::string, Node *> &nodes) override;
+};
+
+void FoldShapeSliceConcatTileFuser::BuildPattern() {
+  // Create patterns
+  auto shape_in =
+      CreatePattern("shape_in")->IsOperationInputOperand(NNADAPTER_SHAPE, 0);
+  auto shape_dtype = CreatePattern("shape_dtype")
+                         ->IsOperationInputOperand(NNADAPTER_SHAPE, 1)
+                         ->IsIntermediate();
+  auto shape = CreatePattern("shape", NNADAPTER_SHAPE)->IsIntermediate();
+  auto shape_out = CreatePattern("shape_out")
+                       ->IsOperationOutputOperand(NNADAPTER_SHAPE, 0)
+                       ->IsOperationInputOperand(NNADAPTER_SLICE, 0)
+                       ->IsIntermediate();
+
+  auto slice_axes = CreatePattern("slice_axes")
+                        ->IsOperationInputOperand(NNADAPTER_SLICE, 1)
+                        ->IsIntermediate();
+  auto slice_starts = CreatePattern("slice_starts")
+                          ->IsOperationInputOperand(NNADAPTER_SLICE, 2)
+                          ->IsIntermediate();
+  auto slice_ends = CreatePattern("slice_ends")
+                        ->IsOperationInputOperand(NNADAPTER_SLICE, 3)
+                        ->IsIntermediate();
+  auto slice_steps = CreatePattern("slice_steps")
+                         ->IsOperationInputOperand(NNADAPTER_SLICE, 4)
+                         ->IsIntermediate();
+  auto slice = CreatePattern("slice", NNADAPTER_SLICE)->IsIntermediate();
+  auto slice_out = CreatePattern("slice_out")
+                       ->IsOperationOutputOperand(NNADAPTER_SLICE, 0)
+                       ->IsOperationInputOperand(NNADAPTER_CONCAT, 1)
+                       ->IsIntermediate();
+
+  auto concat_in0 = CreatePattern("concat_in0")
+                        ->IsConstantOperand()
+                        ->IsOperationInputOperand(NNADAPTER_CONCAT, 0)
+                        ->IsIntermediate();
+  auto concat_in2 = CreatePattern("concat_in2")
+                        ->IsConstantOperand()
+                        ->IsOperationInputOperand(NNADAPTER_CONCAT, 2)
+                        ->IsIntermediate();
+  auto concat_axis = CreatePattern("concat_axis")
+                         ->IsOperationInputOperand(NNADAPTER_CONCAT, 3)
+                         ->IsIntermediate();
+  auto concat = CreatePattern("concat", NNADAPTER_CONCAT)->IsIntermediate();
+  auto concat_out = CreatePattern("concat_out")
+                        ->IsOperationOutputOperand(NNADAPTER_CONCAT, 0)
+                        ->IsOperationInputOperand(NNADAPTER_TILE, 1)
+                        ->IsIntermediate();
+
+  auto tile = CreatePattern("tile", NNADAPTER_TILE);
+
+  // Create the topological connections for the above patterns
+  std::vector<Pattern *> shape_ins{shape_in, shape_dtype};
+  shape_ins >> *shape >> *shape_out;
+  std::vector<Pattern *> slice_ins{
+      shape_out, slice_axes, slice_starts, slice_ends, slice_steps};
+  slice_ins >> *slice >> *slice_out;
+  std::vector<Pattern *> concat_ins{
+      concat_in0, slice_out, concat_in2, concat_axis};
+  concat_ins >> *concat >> *concat_out >> *tile;
+}
+
+bool FoldShapeSliceConcatTileFuser::HandleMatchedResults(
+    core::Model *model, const std::map<std::string, Node *> &nodes) {
+  auto shape_in_data = nodes.at("shape_in")->operand->type.dimensions.data;
+  auto start_index = *reinterpret_cast<int32_t *>(
+      nodes.at("slice")->operation->input_operands[2]->buffer);
+  std::vector<int32_t> repeat_times{
+      *reinterpret_cast<int32_t *>(nodes.at("concat_in0")->operand->buffer),
+      shape_in_data[start_index],
+      *reinterpret_cast<int32_t *>(nodes.at("concat_in2")->operand->buffer),
+  };
+  std::vector<int32_t> dims{3};
+  auto tile_in1 = AddInt32ConstantOperand(model, repeat_times.data(), dims);
+  nodes.at("tile")->operation->input_operands[1] = tile_in1;
+  return true;
+}
+
+NNADAPTER_EXPORT void FoldShapeSliceConcatTile(core::Model *model) {
+  NNADAPTER_VLOG(5) << "Apply FoldShapeSliceConcatTileFuser";
+  bool stop;
+  do {
+    FoldShapeSliceConcatTileFuser fold_shape_slice_concat_tile_fuser;
+    stop = fold_shape_slice_concat_tile_fuser.Apply(model) == 0;
+  } while (!stop);
+}
+
 }  // namespace nnadapter

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

@@ -1108,6 +1108,31 @@ void NCHW2NHWCDataLayoutConverter::ConvertStack(core::Operation* operation) {
   }
 }
 
+void NCHW2NHWCDataLayoutConverter::ConvertTile(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
+  NNADAPTER_CHECK(IsConstantOperand(input_operands[1]));
+  int32_t* repeat_data = reinterpret_cast<int32_t*>(input_operands[1]->buffer);
+  int32_t repeat_count = input_operands[1]->length / sizeof(int32_t);
+  std::vector<int32_t> repeat(repeat_count);
+  auto input_permutation = GetPermutation(input_operand);
+  NNADAPTER_CHECK_EQ(repeat_count,
+                     static_cast<int32_t>(input_permutation.size()));
+  for (int i = 0; i < repeat_count; i++) {
+    repeat[i] = repeat_data[input_permutation[i]];
+  }
+  memcpy(repeat_data, repeat.data(), input_operands[1]->length);
+  TransposeOperand(output_operand, input_permutation);
+  SetPermutation(output_operand, input_permutation);
+}
+
 void NCHW2NHWCDataLayoutConverter::ConvertTranspose(
     core::Operation* operation) {
   auto& input_operands = operation->input_operands;
@@ -1295,6 +1320,9 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
       case NNADAPTER_STACK:
         ConvertStack(operation);
         break;
+      case NNADAPTER_TILE:
+        ConvertTile(operation);
+        break;
       case NNADAPTER_TRANSPOSE:
         ConvertTranspose(operation);
         break;

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

@@ -123,6 +123,7 @@ NNADAPTER_EXPORT void ConvertQuantizationSymmToAsymm(core::Model* model) {
       case NNADAPTER_SQUEEZE:
       case NNADAPTER_SWISH:
       case NNADAPTER_TANH:
+      case NNADAPTER_TILE:
       case NNADAPTER_TRANSPOSE:
       case NNADAPTER_UNSQUEEZE: {
         ConvertOperandSymmToAsymm(input_operands[0], 128);

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

@@ -181,7 +181,8 @@ PatternMatcher::Pattern::IsOperationOutputOperand(NNAdapterOperationType type,
 NNADAPTER_EXPORT PatternMatcher::Pattern *PatternMatcher::Pattern::IsOperation(
     NNAdapterOperationType type) {
   conditions.emplace_back([type](const Node *node) {
-    return node && node->IsOperation() && node->operation->type == type;
+    return node && node->IsOperation() &&
+           (node->operation->type == type || type == NNADAPTER_UNKNOWN);
   });
   return this;
 }
@@ -214,8 +215,11 @@ NNADAPTER_EXPORT size_t PatternMatcher::Apply(core::Model *model) {
     return 0;
   }
   auto subgraphs = DetectPatterns();
+  NNADAPTER_VLOG(5) << subgraphs.size() << " subgraphs detected!";
   UniquePatterns(&subgraphs);
+  NNADAPTER_VLOG(5) << subgraphs.size() << " subgraphs unique!";
   ValidatePatterns(&subgraphs);
+  NNADAPTER_VLOG(5) << subgraphs.size() << " subgraphs valid!";
   RemoveOverlappedPatterns(&subgraphs);
   NNADAPTER_VLOG(5) << subgraphs.size() << " subgraphs matched!";
   // Notify to handle the matched subgraphs, and collect the intermediate