Merge branch 'develop' of https://github.com/PaddlePaddle/paddle-mobile into dev-latest

src/common/types.cpp

@@ -34,6 +34,7 @@ const char *G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
 const char *G_OP_TYPE_LRN = "lrn";
 const char *G_OP_TYPE_MUL = "mul";
 const char *G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
+const char *G_OP_TYPE_POLYGON_BOX_TRANSFORM = "polygon_box_transform";
 const char *G_OP_TYPE_POOL2D = "pool2d";
 const char *G_OP_TYPE_PRIOR_BOX = "prior_box";
 const char *G_OP_TYPE_RELU = "relu";
@@ -94,6 +95,7 @@ std::unordered_map<
         {G_OP_TYPE_FUSION_CONV_BN_ADD_RELU, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_PRIOR_BOX, {{"Image", "Input"}, {"Boxes", "Variances"}}},
         {G_OP_TYPE_MULTICLASS_NMS, {{"BBoxes", "Scores"}, {"Out"}}},
+        {G_OP_TYPE_POLYGON_BOX_TRANSFORM, {{"Input"}, {"Output"}}},
         {G_OP_TYPE_FC, {{"X", "Y", "Z"}, {"Out"}}},
         {G_OP_TYPE_RESHAPE, {{"X"}, {"Out"}}},
         {G_OP_TYPE_DEPTHWISE_CONV, {{"Input"}, {"Output"}}},

src/fpga/bias_scale.cpp

@@ -27,9 +27,6 @@ void align_element(float **data_in, int num_per_div_before_alignment, int num) {
       (num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
   int num_per_div_after_alignment =
       align_to_x(num_per_div_before_alignment, BS_NUM_ALIGNMENT);
-  if (num_per_div_before_alignment == num_per_div_after_alignment) {
-    return;
-  }
   int num_element =
       2 * div_num * num_per_div_after_alignment;  // including bias & scale
   float *ptr_aligned =

src/framework/load_ops.h

@@ -199,6 +199,9 @@ LOAD_OP3(pool2d, CPU, MALI_GPU, FPGA);
 #ifdef MULTICLASSNMS_OP
 LOAD_OP1(multiclass_nms, CPU);
 #endif
+#ifdef POLYGONBOXTRANSFORM_OP
+LOAD_OP1(polygon_box_transform, CPU);
+#endif
 #ifdef SUM_OP
 LOAD_OP1(sum, CPU);
 #endif

src/operators/kernel/arm/polygon_box_transform_kernel.cpp

+/* Copyright (c) 2018 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. */
+
+#ifdef POLYGONBOXTRANSFORM_OP
+
+#include "operators/kernel/polygon_box_transform_kernel.h"
+#include "operators/kernel/central-arm-func/polygon_box_transform_arm_func.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <>
+bool PolygonBoxTransformKernel<CPU, float>::Init(
+    PolygonBoxTransformParam<CPU> *param) {
+  return true;
+}
+
+template <>
+void PolygonBoxTransformKernel<CPU, float>::Compute(
+    const PolygonBoxTransformParam<CPU> &param) const {
+  PolygonBoxTransformCompute<float>(param);
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/kernel/central-arm-func/polygon_box_transform_arm_func.h

+/* Copyright (c) 2018 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. */
+
+#ifdef POLYGONBOXTRANSFORM_OP
+#pragma once
+
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename P>
+void PolygonBoxTransformCompute(const PolygonBoxTransformParam<CPU>& param) {
+  const auto* input = param.Input();
+  const auto& input_dims = input->dims();
+  const auto* input_data = input->data<float>();
+  auto* output = param.Output();
+  auto* output_data = output->mutable_data<float>();
+
+  int64_t batch_size = input_dims[0];
+  int64_t geo_channel = input_dims[1];
+  int64_t height = input_dims[2];
+  int64_t width = input_dims[3];
+  int64_t id = 0;
+  for (int64_t id_n = 0; id_n < batch_size * geo_channel; ++id_n) {
+    for (int64_t id_h = 0; id_h < height; ++id_h) {
+      for (int64_t id_w = 0; id_w < width; ++id_w) {
+        id = id_n * height * width + width * id_h + id_w;
+        if (id_n % 2 == 0) {
+          output_data[id] = id_w * 4 - input_data[id];
+        } else {
+          output_data[id] = id_h * 4 - input_data[id];
+        }
+      }
+    }
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/kernel/polygon_box_transform_kernel.h

+/* Copyright (c) 2018 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. */
+
+#ifdef POLYGONBOXTRANSFORM_OP
+
+#pragma once
+
+#include "framework/operator.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename DeviceType, typename T>
+class PolygonBoxTransformKernel
+    : public framework::OpKernelBase<DeviceType,
+                                     PolygonBoxTransformParam<DeviceType>> {
+ public:
+  void Compute(const PolygonBoxTransformParam<DeviceType>& param) const;
+  bool Init(PolygonBoxTransformParam<DeviceType>* param);
+};
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/math/gemm.cpp

@@ -3379,7 +3379,7 @@ void Gemm::SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A,
     // 对 B 分块
     NC = L1 / (KC * sizeof(float));
     if (NC == 0) {
-      NC == NR;
+      NC = NR;
     } else {
       int nblock_num = (n + NC - 1) / NC;
       NC = (n + nblock_num - 1) / nblock_num;

src/operators/math/gemm.h

@@ -22,9 +22,11 @@ limitations under the License. */
 #define C(i, j) C[(i)*ldc + (j)]
 
 #if __aarch64__
+#define MR_INT8 4
 #define MR 6
 #define NR 16
 #else
+#define MR_INT8 4
 #define MR 6
 #define NR 8
 #endif
@@ -189,6 +191,8 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
 
   // 8 bits function cluster begins
   // 8 bits int small block inner product
+  void AddDot4x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
+                 int32_t ldc);
   void AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
                  int32_t ldc);
 
@@ -199,6 +203,8 @@ void PackMatrixB(int k, int n, int n_tail, const float *B, int ldb,
                            int8_t *bias);
 
   // 8 bits int pack function
+  void PackMatrixA_4r(int32_t m, int32_t k, int32_t m_tail, const int8_t *A,
+                      int32_t lda, int8_t *buffer);
   void PackMatrixA_6r(int32_t m, int32_t k, int32_t m_tail, const int8_t *A,
                       int32_t lda, int8_t *buffer);
   void PackMatrixB_8c(int32_t k, int32_t n, int32_t n_tail, const int8_t *B,

src/operators/op_param.h

@@ -999,6 +999,28 @@ class MultiClassNMSParam : public OpParam {
 };
 #endif
 
+#ifdef POLYGONBOXTRANSFORM_OP
+template <typename Dtype>
+class PolygonBoxTransformParam : public OpParam {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
+
+ public:
+  PolygonBoxTransformParam(const VariableNameMap &inputs,
+                           const VariableNameMap &outputs,
+                           const AttributeMap &attrs, const Scope &scope) {
+    input_ = InputFrom<GType>(inputs, scope);
+    output_ = OutputFrom<GType>(outputs, scope);
+  }
+  const RType *Input() const { return input_; }
+  RType *Output() const { return output_; }
+
+ private:
+  RType *input_;
+  RType *output_;
+};
+#endif
+
 template <typename Dtype>
 class FeedParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
@@ -2272,6 +2294,7 @@ class ShapeParam : public OpParam {
 };
 #endif
 
+#ifdef QUANT_OP
 template <typename Dtype>
 class QuantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
@@ -2311,7 +2334,9 @@ class QuantizeParam : public OpParam {
   // nearest_zero and nearest_even is valid currently
   RoundType round_type_ = ROUND_NEAREST_AWAY_ZERO;
 };
+#endif
 
+#ifdef DEQUANT_OP
 template <typename Dtype>
 class DequantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
@@ -2339,6 +2364,7 @@ class DequantizeParam : public OpParam {
   RType *activation_scale_;
   float weight_scale_;
 };
+#endif
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/polygon_box_transform_op.cpp

+/* Copyright (c) 2018 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. */
+
+#ifdef POLYGONBOXTRANSFORM_OP
+
+#include "operators/polygon_box_transform_op.h"
+namespace paddle_mobile {
+namespace operators {
+
+template <typename Dtype, typename T>
+void PolygonBoxTransformOp<Dtype, T>::InferShape() const {
+  PADDLE_MOBILE_ENFORCE(this->param_.Input() != nullptr,
+                        "Input (Input) of get_shape op should not be null.");
+  PADDLE_MOBILE_ENFORCE(this->param_.Output() != nullptr,
+                        "Output (Output) of get_shape op should not be null.");
+
+  auto input_dims = this->param_.Input()->dims();
+
+  PADDLE_MOBILE_ENFORCE(input_dims.size() == 4, "input's rank must be 4.");
+  PADDLE_MOBILE_ENFORCE(input_dims[1] % 2 == 0,
+                        "input's second dimension must be even.");
+
+  this->param_.Output()->Resize(input_dims);
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+REGISTER_OPERATOR_CPU(polygon_box_transform, ops::PolygonBoxTransformOp);
+#endif
+
+#endif

src/operators/polygon_box_transform_op.h

+/* Copyright (c) 2018 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. */
+
+#ifdef POLYGONBOXTRANSFORM_OP
+
+#pragma once
+
+#include <string>
+
+#include "framework/operator.h"
+#include "operators/kernel/polygon_box_transform_kernel.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+using paddle_mobile::framework::Tensor;
+
+template <typename DeviceType, typename T>
+class PolygonBoxTransformOp
+    : public framework::OperatorWithKernel<
+          DeviceType, PolygonBoxTransformParam<DeviceType>,
+          operators::PolygonBoxTransformKernel<DeviceType, T>> {
+ public:
+  PolygonBoxTransformOp(const std::string &type, const VariableNameMap &inputs,
+                        const VariableNameMap &outputs,
+                        const framework::AttributeMap &attrs,
+                        std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<
+            DeviceType, PolygonBoxTransformParam<DeviceType>,
+            operators::PolygonBoxTransformKernel<DeviceType, T>>(
+            type, inputs, outputs, attrs, scope) {}
+
+  using framework::OperatorWithKernel<
+      DeviceType, PolygonBoxTransformParam<DeviceType>,
+      operators::PolygonBoxTransformKernel<DeviceType, T>>::OperatorWithKernel;
+  void InferShape() const override;
+
+ protected:
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

test/CMakeLists.txt

@@ -181,6 +181,10 @@ if (NOT FOUND_MATCH)
     ADD_EXECUTABLE(test-multiclassnms-op operators/test_multiclass_nms_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-multiclassnms-op paddle-mobile)
 
+    # gen test
+    ADD_EXECUTABLE(test-polygon-box-transform-op operators/test_polygon_box_transform_op.cpp test_helper.h test_include.h)
+    target_link_libraries(test-polygon-box-transform-op paddle-mobile)
+
     # gen test
     ADD_EXECUTABLE(test-reshape-op operators/test_reshape_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-reshape-op paddle-mobile)

test/operators/test_mul_op.cpp

@@ -12,7 +12,6 @@ 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. */
 
-#include <stdint-gcc.h>
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/mul_op.h"

test/operators/test_polygon_box_transform_op.cpp

+/* Copyright (c) 2018 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 "../test_include.h"
+#include "operators/polygon_box_transform_op.h"
+
+namespace paddle_mobile {
+namespace framework {
+
+template <typename Dtype>
+class TestPolygonBoxTransformOp {
+ public:
+  explicit TestPolygonBoxTransformOp(const Program<Dtype> p) : program_(p) {
+    if (use_optimize_) {
+      to_predict_program_ = program_.optimizeProgram;
+    } else {
+      to_predict_program_ = program_.originProgram;
+    }
+    const std::vector<std::shared_ptr<BlockDesc>> blocks =
+        to_predict_program_->Blocks();
+    for (auto block_desc : blocks) {
+      std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
+      for (auto op : ops) {
+        if (op->Type() == "polygon_box_transform") {
+          DLOG << " attr size: " << op->GetAttrMap().size();
+          DLOG << " inputs size: " << op->GetInputs().size();
+          DLOG << " input is : " << op->Input("Input")[0];
+          input_var_name = op->Input("Input")[0];
+          DLOG << " outputs size: " << op->GetOutputs().size();
+          DLOG << " output is : " << op->Output("Output")[0];
+          output_var_name = op->Output("Output")[0];
+          std::shared_ptr<operators::PolygonBoxTransformOp<Dtype, float>>
+              op_ptr = std::make_shared<
+                  operators::PolygonBoxTransformOp<Dtype, float>>(
+                  op->Type(), op->GetInputs(), op->GetOutputs(),
+                  op->GetAttrMap(), program_.scope);
+          ops_of_block_[*block_desc.get()].push_back(op_ptr);
+          return;
+        }
+      }
+    }
+  }
+
+  std::shared_ptr<Tensor> predict(const Tensor &t) {
+    auto scope = program_.scope;
+    Variable *input_feed_value = scope->Var(input_var_name);
+    auto tensor_input = input_feed_value->GetMutable<LoDTensor>();
+    tensor_input->ShareDataWith(t);
+
+    Variable *output = scope->Var(output_var_name);
+    auto *output_tensor = output->GetMutable<LoDTensor>();
+
+    std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
+    out_tensor.reset(output_tensor);
+
+    predict(t, 0);
+
+    return out_tensor;
+  }
+
+ private:
+  const framework::Program<Dtype> program_;
+  std::shared_ptr<ProgramDesc> to_predict_program_;
+  std::map<framework::BlockDesc,
+           std::vector<std::shared_ptr<OperatorBase<Dtype>>>>
+      ops_of_block_;
+  bool use_optimize_ = false;
+  string input_var_name;
+  string output_var_name;
+
+  void predict(const Tensor &t, int block_id) {
+    std::shared_ptr<BlockDesc> to_predict_block =
+        to_predict_program_->Block(block_id);
+    for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
+      auto op = ops_of_block_[*to_predict_block.get()][j];
+      op->Run();
+    }
+  }
+};
+
+template class TestPolygonBoxTransformOp<CPU>;
+}  // namespace framework
+}  // namespace paddle_mobile
+
+int main() {
+  DLOG << "----------**********----------";
+  DLOG << "begin to run PolygonBoxTransform Test";
+  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  auto program = loader.Load(std::string(g_ocr));
+
+  paddle_mobile::framework::Tensor input;
+  SetupTensor<float>(&input, {1, 8, 1, 2}, static_cast<float>(0),
+                     static_cast<float>(1));
+  auto *input_ptr = input.data<float>();
+  for (int i = 0; i < 16; ++i) {
+    *(input_ptr + i) = i;
+  }
+  DLOG << "input : ";
+  for (int i = 0; i < input.numel(); ++i) {
+    DLOG << " index " << i << " : " << input_ptr[i];
+  }
+
+  paddle_mobile::framework::TestPolygonBoxTransformOp<paddle_mobile::CPU>
+      testPolygonBoxTransformOp(program);
+
+  auto output = testPolygonBoxTransformOp.predict(input);
+  auto *output_ptr = output->data<float>();
+
+  DLOG << "output : ";
+  for (int i = 0; i < output->numel(); ++i) {
+    DLOG << " index " << i << " : " << output_ptr[i];
+  }
+  return 0;
+}

tools/op.cmake

@@ -195,6 +195,7 @@ if(NOT FOUND_MATCH)
   set(LRN_OP ON)
   set(MUL_OP ON)
   set(MULTICLASSNMS_OP ON)
+  set(POLYGONBOXTRANSFORM_OP ON)
   set(POOL_OP ON)
   set(PRIORBOX_OP ON)
   set(RELU_OP ON)
@@ -238,6 +239,7 @@ endif()
   # option(LRN_OP "" ON)
   # option(MUL_OP "" ON)
   # option(MULTICLASSNMS_OP "" ON)
+  # option(POLYGONBOXTRANSFORM_OP "" ON)
   # option(POOL_OP "" ON)
   # option(PRIORBOX_OP "" ON)
   # option(RELU_OP "" ON)
@@ -292,6 +294,9 @@ endif()
 if (MULTICLASSNMS_OP)
   add_definitions(-DMULTICLASSNMS_OP)
 endif()
+if (POLYGONBOXTRANSFORM_OP)
+  add_definitions(-DPOLYGONBOXTRANSFORM_OP)
+endif()
 if (POOL_OP)
   add_definitions(-DPOOL_OP)
 endif()