[luci] infer dynamic shape for pad

This infers dynamic shape for pad.

DCO-1.0-Signed-off-by: JuYoung Lee <[email protected]>

compiler/luci/export/include/luci/CircleExporter.h

@@ -36,14 +36,9 @@ class CircleExporter
     virtual ~Contract() = default;
 
   public: // Client -> Exporter
-    // Input Graph (to be exported)
-    // Exporter expects a loco graph that consists of Circle nodes
-    virtual loco::Graph *graph(void) const = 0;
-
     // Input Module (to be exported)
     // Exporter expects a luci module that consists of loco graphs
-    // TODO make this pure virtual
-    virtual luci::Module *module(void) const;
+    virtual luci::Module *module(void) const = 0;
 
   public: // Exporter -> Client
     // Exporter calls store for export data

compiler/luci/export/include/luci/CircleFileExpContract.h

@@ -40,7 +40,6 @@ struct CircleFileExpContract : public luci::CircleExporter::Contract
   virtual ~CircleFileExpContract() = default;
 
 public:
-  loco::Graph *graph(void) const final { return nullptr; }
   luci::Module *module(void) const final { return _module; }
 
 public:

compiler/luci/export/src/CircleExporter.cpp

@@ -26,9 +26,6 @@
 namespace luci
 {
 
-// TODO remove this
-Module *CircleExporter::Contract::module(void) const { return nullptr; }
-
 CircleExporter::CircleExporter()
 {
   // NOTHING TO DO
@@ -48,17 +45,10 @@ bool CircleExporter::invoke(Contract *contract) const
     return contract->store(ptr, size);
   }
 
-  auto graph = contract->graph();
-  if (graph == nullptr)
-    return false;
-
-  CircleExporterImpl impl(graph);
-
-  const char *ptr = impl.getBufferPointer();
-  const size_t size = impl.getBufferSize();
+  // NOTE some unit tests calls with nullptr module, cannot add assert here
+  // TODO fix those unit tests and add assert(false)
 
-  // we just send one time
-  return contract->store(ptr, size);
+  return false;
 }
 
 } // namespace luci

compiler/luci/export/src/CircleExporter.test.cpp

@@ -33,12 +33,12 @@ class SampleGraphContract : public luci::CircleExporter::Contract
   SampleGraphContract() : luci::CircleExporter::Contract(), _buffer(new std::vector<char>)
   {
     // create needed entities
-    _g = loco::make_graph();
-    auto graph_input = _g->inputs()->create();
-    auto graph_output = _g->outputs()->create();
-    input_node = _g->nodes()->create<luci::CircleInput>();
-    output_node = _g->nodes()->create<luci::CircleOutput>();
-    relu_node = _g->nodes()->create<luci::CircleRelu>();
+    auto g = loco::make_graph();
+    auto graph_input = g->inputs()->create();
+    auto graph_output = g->outputs()->create();
+    input_node = g->nodes()->create<luci::CircleInput>();
+    output_node = g->nodes()->create<luci::CircleOutput>();
+    relu_node = g->nodes()->create<luci::CircleRelu>();
 
     // link nodes and link them to graph
     relu_node->features(input_node);
@@ -57,9 +57,12 @@ class SampleGraphContract : public luci::CircleExporter::Contract
 
     graph_output->shape({1, 2, 3, 4});
     graph_output->dtype(loco::DataType::FLOAT32);
+
+    _m = std::unique_ptr<luci::Module>{new luci::Module};
+    _m->add(std::move(g));
   }
 
-  loco::Graph *graph(void) const override { return _g.get(); }
+  luci::Module *module(void) const override { return _m.get(); }
 
 public:
   bool store(const char *ptr, const size_t size) const override
@@ -77,7 +80,7 @@ class SampleGraphContract : public luci::CircleExporter::Contract
   luci::CircleRelu *relu_node;
 
 private:
-  std::unique_ptr<loco::Graph> _g;
+  std::unique_ptr<luci::Module> _m;
   std::unique_ptr<std::vector<char>> _buffer;
 };
 

compiler/luci/export/src/CircleExporterImpl.cpp

@@ -133,7 +133,6 @@ namespace luci
 using namespace circle;
 using namespace flatbuffers;
 
-CircleExporterImpl::CircleExporterImpl(loco::Graph *graph) { exportGraph(graph); }
 CircleExporterImpl::CircleExporterImpl(Module *module) { exportModule(module); }
 
 ::flatbuffers::Offset<::circle::SubGraph>
@@ -148,61 +147,6 @@ CircleExporterImpl::exportSubgraph(SerializedGraphData &gd)
   return subgraph;
 }
 
-void CircleExporterImpl::exportGraph(loco::Graph *graph)
-{
-  // do graph optimization
-  optimize(graph);
-
-  _builder.Clear();
-
-  SerializedModelData md;
-  SerializedGraphData gd;
-
-  // This version is taken from comment in fbs
-  constexpr uint32_t version = 0;
-
-  // set Subgraph name
-  gd._name = graph->name();
-
-  // TODO set this value properly
-  gd._data_format = circle::DataFormat::DataFormat_CHANNELS_LAST;
-
-  // prepare model data
-  prepareModelData(_builder, md);
-
-  // parse graph into SerializedModelData structure
-  exportOpDefinedTensors(graph, _builder, md, gd);
-
-  // NOTE Invoke these register functions only after each node is annotated with its tensor_index
-  registerGraphInputTensors(graph, gd);
-  registerGraphOutputTensors(graph, gd);
-
-  exportNodes(graph, _builder, md, gd);
-
-  // encode operator codes
-  auto operator_codes = encodeOperatorCodes(_builder, md._operator_codes);
-
-  // Subgraphs
-  Offset<SubGraph> subgraph = exportSubgraph(gd);
-  auto subgraphs = _builder.CreateVector(std::vector<Offset<SubGraph>>{subgraph});
-
-  // Description
-  std::string description_str = "nnpackage";
-  auto description = _builder.CreateString(description_str);
-
-  // Metadata
-  auto metadata_vec = createCircleMetadataVector(_builder, md);
-  auto metadata = _builder.CreateVector(std::vector<Offset<Metadata>>(metadata_vec));
-
-  // create array of buffers
-  auto buffers = _builder.CreateVector(md._buffers);
-
-  // Model
-  auto model_offset = CreateModel(_builder, version, operator_codes, subgraphs, description,
-                                  buffers, 0 /* metadata_buffer */, metadata);
-  FinishModelBuffer(_builder, model_offset);
-}
-
 void CircleExporterImpl::exportModule(Module *module)
 {
   assert(module->size() > 0);
@@ -248,7 +192,7 @@ void CircleExporterImpl::exportModule(Module *module)
   auto operator_codes = encodeOperatorCodes(_builder, md._operator_codes);
 
   // Description
-  std::string description_str = "nnpackage";
+  std::string description_str = "ONE-luci/export";
   auto description = _builder.CreateString(description_str);
 
   // Metadata

compiler/luci/export/src/CircleExporterImpl.h

@@ -38,7 +38,6 @@ class CircleExporterImpl
   CircleExporterImpl() = delete;
   ~CircleExporterImpl() = default;
 
-  explicit CircleExporterImpl(loco::Graph *graph);
   explicit CircleExporterImpl(Module *module);
 
   /**
@@ -59,12 +58,6 @@ class CircleExporterImpl
    */
   flatbuffers::Offset<circle::SubGraph> exportSubgraph(SerializedGraphData &gd);
 
-  /**
-   * @brief root function that writes graph into internal buffer
-   * @param graph
-   */
-  void exportGraph(loco::Graph *graph);
-
   /**
    * @brief root function that writes Module into internal buffer
    * @param module

compiler/luci/service/src/CircleShapeInferenceRule.cpp

@@ -237,27 +237,33 @@ loco::NodeShape use_paddings(const CIRCLENODE *node, const luci::CircleConst *pa
   {
     int32_t idx = ni * 2;
     int value = input_shape.dim(ni).value();
-    if (paddings->dtype() == S32)
+    if (!input_shape.dim(ni).known())
     {
-      value += paddings->at<S32>(idx + 0); // left
-      value += paddings->at<S32>(idx + 1); // right
+      output_shape.dim(ni).unset();
     }
     else
     {
-      auto pl = paddings->at<S64>(idx + 0);
-      auto pr = paddings->at<S64>(idx + 1);
-      auto max = static_cast<int64_t>(std::numeric_limits<int32_t>::max());
-      auto low = static_cast<int64_t>(std::numeric_limits<int32_t>::lowest());
-      LUCI_ASSERT(pl <= max, "paddings is over 32 bit limit");
-      LUCI_ASSERT(pl >= low, "paddings is over 32 bit limit");
-      LUCI_ASSERT(pr <= max, "paddings is over 32 bit limit");
-      LUCI_ASSERT(pr >= low, "paddings is over 32 bit limit");
-      value += static_cast<int32_t>(pl); // left
-      value += static_cast<int32_t>(pr); // right
+      if (paddings->dtype() == S32)
+      {
+        value += paddings->at<S32>(idx + 0); // left
+        value += paddings->at<S32>(idx + 1); // right
+      }
+      else
+      {
+        auto pl = paddings->at<S64>(idx + 0);
+        auto pr = paddings->at<S64>(idx + 1);
+        auto max = static_cast<int64_t>(std::numeric_limits<int32_t>::max());
+        auto low = static_cast<int64_t>(std::numeric_limits<int32_t>::lowest());
+        LUCI_ASSERT(pl <= max, "paddings is over 32 bit limit");
+        LUCI_ASSERT(pl >= low, "paddings is over 32 bit limit");
+        LUCI_ASSERT(pr <= max, "paddings is over 32 bit limit");
+        LUCI_ASSERT(pr >= low, "paddings is over 32 bit limit");
+        value += static_cast<int32_t>(pl); // left
+        value += static_cast<int32_t>(pr); // right
+      }
+      output_shape.dim(ni) = value;
     }
-    output_shape.dim(ni) = value;
   }
-
   return loco::NodeShape{output_shape};
 }
 
@@ -2599,7 +2605,8 @@ bool CircleShapeInferenceRule::infer(const loco::Node *node, loco::NodeShape &sh
     else
       shape = circle_node->accept(&alg);
   }
-
+  VERBOSE(l, 1) << "[icodo] shape status: " << is_shape_undefined << ", " << is_shape_none << ", "
+                << is_scalar;
   VERBOSE(l, 1) << "[luci] shape: " << circle_node->name();
   VERBOSE(l, 1) << "              own_shape: " << own_shape(circle_node)
                 << " -> infer: " << shape.as<loco::TensorShape>();

compiler/luci/service/src/Nodes/CirclePad.test.cpp

@@ -16,6 +16,8 @@
 
 #include "luci/Service/CircleNodeClone.h"
 
+#include "luci/Service/CircleShapeInference.h"
+
 #include <gtest/gtest.h>
 
 TEST(CloneNodeTest, clone_Pad)
@@ -31,3 +33,42 @@ TEST(CloneNodeTest, clone_Pad)
   auto cloned_pad = dynamic_cast<luci::CirclePad *>(cloned);
   ASSERT_NE(nullptr, cloned_pad);
 }
+
+TEST(ShapeRuleTest, padding_dynamic_shape)
+{
+  luci::CirclePad pad;
+  luci::CircleInput input;
+  // Use circle input as paddings
+  luci::CircleConst padddings;
+
+  loco::TensorShape shape;
+  luci::sinf::Rule shape_inf_rule;
+
+  input.shape({1, 2, 3, 4});
+  input.shape_status(luci::ShapeStatus::VALID);
+  input.dim(2).unset();
+
+  padddings.dtype(loco::DataType::S64);
+  padddings.shape({4, 2});
+  padddings.shape_status(luci::ShapeStatus::VALID);
+
+  const loco::DataType S64 = loco::DataType::S64;
+  uint32_t t = 64 * 8;
+  padddings.size<S64>(t);
+
+  pad.input(&input);
+  pad.paddings(&padddings);
+
+  ASSERT_TRUE(shape_inf_rule.infer(&pad, shape));
+  ASSERT_EQ(shape.rank(), 4);
+  ASSERT_TRUE(shape.dim(0).known());
+  ASSERT_TRUE(shape.dim(1).known());
+  ASSERT_FALSE(shape.dim(2).known());
+  ASSERT_TRUE(shape.dim(3).known());
+
+  ASSERT_EQ(1, shape.dim(0).value());
+  ASSERT_EQ(2, shape.dim(1).value());
+  ASSERT_EQ(0, shape.dim(2).value());
+  ASSERT_EQ(4, shape.dim(3).value());
+  pad.drop();
+}

compiler/luci/tester/src/WriteTester.cpp

@@ -42,11 +42,6 @@ void show_error_message(const char *progname, std::ostream &os, const std::strin
 struct CircleExpContract : public luci::CircleExporter::Contract
 {
 public:
-  CircleExpContract(loco::Graph *graph, const std::string &filename)
-    : _graph(graph), _filepath(filename)
-  {
-    // NOTHING TO DO
-  }
   CircleExpContract(luci::Module *module, const std::string &filename)
     : _module(module), _filepath(filename)
   {
@@ -55,15 +50,12 @@ struct CircleExpContract : public luci::CircleExporter::Contract
   virtual ~CircleExpContract() = default;
 
 public:
-  loco::Graph *graph(void) const final { return _graph; }
-
   luci::Module *module(void) const final { return _module; }
 
 public:
   bool store(const char *ptr, const size_t size) const final;
 
 private:
-  loco::Graph *_graph{nullptr};
   luci::Module *_module{nullptr};
   const std::string _filepath;
 };

onert-micro/onert-micro/include/execute/kernels/ComparisonCommon.h

@@ -110,8 +110,8 @@ template <typename T> void evalComparisonGeneric(OMRuntimeKernel *runtime_kernel
   }
 }
 
-template <typename T>
-void evalQuantizedComparisonGeneric(OMRuntimeKernel *runtime_kernel, bool F(T, T))
+template <typename T, typename AccType>
+void evalQuantizedComparisonGeneric(OMRuntimeKernel *runtime_kernel, bool F(AccType, AccType))
 {
   const circle::Tensor *input1 = nullptr;
   const circle::Tensor *input2 = nullptr;

onert-micro/onert-micro/include/pal/common/PALComparisons.h

@@ -71,11 +71,11 @@ inline void ComparisonNoScaling(const int64_t flat_size, const T *input1_data, c
   }
 }
 
-template <typename T>
+template <typename T, typename AccType>
 inline void BroadcastComparison4DSlowWithScaling(
   const core::ComparisonParams &op_params, const core::OMRuntimeShape &unextended_input1_shape,
   const T *input1_data, const core::OMRuntimeShape &unextended_input2_shape, const T *input2_data,
-  const core::OMRuntimeShape &unextended_output_shape, bool *output_data, bool F(T, T))
+  const core::OMRuntimeShape &unextended_output_shape, bool *output_data, bool F(AccType, AccType))
 {
   const BroadcastComparison4DSlowCommon dims = BroadcastComparison4DSlowPreprocess(
     unextended_input1_shape, unextended_input2_shape, unextended_output_shape);
@@ -118,10 +118,10 @@ inline void BroadcastComparison4DSlowWithScaling(
   }
 }
 
-template <typename T>
+template <typename T, typename AccType>
 inline void ComparisonWithScaling(const core::ComparisonParams &op_params, const int64_t flat_size,
                                   const T *input1_data, const T *input2_data, bool *output_data,
-                                  bool F(T, T))
+                                  bool F(AccType, AccType))
 {
   int left_shift = op_params.left_shift;
   int32_t input1_offset = op_params.input1_offset;