Merge pull request #285 from cocodark/develop

fix #284

src/framework/tensor.h

@@ -132,13 +132,6 @@ class Tensor {
         reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
   }
 
-  inline void *mutable_data() {
-    //  PADDLE_ENFORCE(this->holder_ != nullptr,
-    //                 "Cannot invoke mutable data if current hold
-    //                 nothing.");
-    return mutable_data(holder_->type());
-  }
-
   /**
    * @brief     Return a pointer to mutable memory block.
    *

src/operators/kernel/arm/sigmoid_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. */
+
+#include "../sigmoid_kernel.h"
+#if __ARM_NEON
+#include "../../math/math_func_neon.h"
+#endif
+
+namespace paddle_mobile {
+namespace operators {
+
+using framework::DDim;
+using framework::Tensor;
+
+void sigmoid(const Tensor *X, Tensor *Y) {
+#if __ARM_NEON
+  DLOG << "step1";
+  const float *input = X->data<float>();
+  DLOG << "step11";
+
+  float *output = Y->mutable_data<float>();
+  DLOG << "step2";
+
+  const DDim &dDim = X->dims();
+  DLOG << "step3";
+
+  int axis_index = 1;
+  if (dDim.size() < 4) {
+    axis_index = 0;
+  }
+  DLOG << "step4";
+
+  DDim outer_ddim =
+      paddle_mobile::framework::slice_ddim(dDim, 0, axis_index + 1);
+  DDim inner_ddim =
+      paddle_mobile::framework::slice_ddim(dDim, axis_index + 1, dDim.size());
+  DLOG << "step5";
+
+  int out_size = paddle_mobile::framework::product(outer_ddim);
+  int inner_size = paddle_mobile::framework::product(inner_ddim);
+  DLOG << "step6";
+
+#pragma omp parallel for
+  DLOG << "outsize=" << out_size;
+  DLOG << "innersize=" << inner_size;
+  for (int i = 0; i < out_size; ++i) {
+    const float *input_outer_ptr = input + i * inner_size;
+    float *output_outer_ptr = output + i * inner_size;
+    int nn = inner_size >> 2;
+    int remain = inner_size - (nn << 2);
+    float32x4_t _one = vdupq_n_f32(1.f);
+    for (; nn > 0; nn--) {
+      float32x4_t data = vld1q_f32(input_outer_ptr);
+      data = vnegq_f32(data);
+      data = exp_ps(data);
+      data = vaddq_f32(data, _one);
+      float32x4_t out_data = vrecpeq_f32(data);
+      out_data = vmulq_f32(vrecpsq_f32(data, out_data), out_data);
+      vst1q_f32(output_outer_ptr, out_data);
+
+      input_outer_ptr += 4;
+      output_outer_ptr += 4;
+    }
+    for (; remain > 0; remain--) {
+      *output_outer_ptr = 1.f / (1.f + exp(-*input_outer_ptr));
+      output_outer_ptr++;
+      input_outer_ptr++;
+    }
+  }
+#endif
+}
+
+template <>
+void SigmoidKernel<CPU, float>::Compute(const SigmoidParam &param) const {
+  const Tensor *in_x = param.InputX();
+  Tensor *out = param.Out();
+  auto x_dims = in_x->dims();
+  out->Resize(x_dims);
+  sigmoid(in_x, out);
+}
+
+template class SigmoidKernel<CPU, float>;
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/kernel/sigmoid_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. */
+
+#pragma once
+
+#include "framework/operator.h"
+#include "operators/op_param.h"
+namespace paddle_mobile {
+namespace operators {
+using framework::OpKernelBase;
+void sigmoid(const Tensor* X, Tensor* Y);
+template <typename DeviceType, typename T>
+class SigmoidKernel : public OpKernelBase<DeviceType, SigmoidParam> {
+ public:
+  void Compute(const SigmoidParam& param) const override;
+};
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/kernel/softmax_kernel.h

@@ -21,6 +21,8 @@ namespace paddle_mobile {
 namespace operators {
 using framework::OpKernelBase;
 
+void simoid(Tensor *X, Tensor *Y);
+
 template <typename DeviceType, typename T>
 class SoftmaxKernel : public OpKernelBase<DeviceType, SoftmaxParam> {
  public:

src/operators/math/softmax.cpp

@@ -11,11 +11,12 @@ 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. */
-
+#define __ARM_NEON true
 #include "operators/math/softmax.h"
 #include "common/types.h"
 #if __ARM_NEON
 #include <math.h>
+#include <algorithm>
 #include "operators/math/math_func_neon.h"
 #endif
 
@@ -108,7 +109,7 @@ class SoftmaxFuntor<CPU, T> {
     // sum exp
     sum(exp_sub_max, sumptr, inner_size, out_size);
     // div
-    auto *out_ptr = static_cast<float *>(Y->mutable_data());
+    auto *out_ptr = Y->mutable_data<float>();
     for (int l = 0; l < out_size; ++l) {
       const float *input_outer_ptr = exp_sub_max + l * inner_size;
       float *output_outer_ptr = out_ptr + l * inner_size;

src/operators/op_param.h

@@ -542,6 +542,22 @@ class SoftmaxParam : public OpParam {
   Tensor *input_x_;
   Tensor *out_;
 };
+
+class SigmoidParam : public OpParam {
+ public:
+  SigmoidParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
+               const framework::AttributeMap &attrs,
+               const framework::Scope &scope) {
+    input_x_ = InputXFrom<framework::Tensor>(inputs, scope);
+    out_ = OutFrom<framework::Tensor>(outputs, scope);
+  }
+  const Tensor *InputX() const { return input_x_; }
+  Tensor *Out() const { return out_; }
+
+ private:
+  Tensor *input_x_;
+  Tensor *out_;
+};
 class MultiClassNMSParam : public OpParam {
  public:
   MultiClassNMSParam(const VariableNameMap &inputs,

src/operators/sigmoid_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. */
+
+#include "operators/sigmoid_op.h"
+
+namespace paddle_mobile {
+namespace operators {
+template <typename DeviceType, typename T>
+void SigmoidOp<DeviceType, T>::InferShape() const {
+  param_.Out()->Resize(param_.InputX()->dims());
+}
+template class SigmoidOp<CPU, float>;
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+USE_OP(sigmoid);
+REGISTER_OPERATOR(sigmoid, ops::SigmoidOp);

src/operators/sigmoid_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. */
+
+#pragma once
+
+#include <framework/operator.h>
+#include <operators/op_param.h>
+#include <string>
+#include "operators/kernel/sigmoid_kernel.h"
+
+namespace paddle_mobile {
+namespace operators {
+template <typename DeviceType, typename T>
+class SigmoidOp : public framework::OperatorWithKernel<DeviceType> {
+ public:
+  SigmoidOp(const std::string &type, const VariableNameMap &inputs,
+            const VariableNameMap &outputs,
+            const framework::AttributeMap &attrs,
+            std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<DeviceType>(type, inputs, outputs, attrs,
+                                                  scope),
+        param_(inputs, outputs, attrs, *scope) {}
+
+  using framework::OperatorWithKernel<DeviceType>::OperatorWithKernel;
+
+  void InferShape() const override;
+
+  void Run() const {
+    operators::SigmoidKernel<DeviceType, T> kernel;
+    kernel.Compute(param_);
+    this->ClearVariables({"X"});
+  }
+
+ private:
+  SigmoidParam param_;
+};
+}  // namespace operators
+}  // namespace paddle_mobile

test/CMakeLists.txt

@@ -82,3 +82,7 @@ target_link_libraries(test-enforce paddle-mobile)
 # gen test
 ADD_EXECUTABLE(test-googlenet net/test_googlenet.cpp test_helper.h  test_include.h executor_for_test.h)
 target_link_libraries(test-googlenet paddle-mobile)
+
+# gen test
+ADD_EXECUTABLE(test-sigmoid operators/test_sigmoid_op.cpp  test_include.h)
+target_link_libraries(test-sigmoid paddle-mobile)

test/executor_for_test.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "operators/conv_op.h"
 #include "operators/pool_op.h"
 #include "operators/reshape_op.h"
+#include "operators/sigmoid_op.h"
 #include "operators/softmax_op.h"
 #include "operators/transpose_op.h"
 

test/operators/test_sigmoid_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. */
+
+#include "../../src/operators/kernel/sigmoid_kernel.h"
+#include "../test_helper.h"
+#include "./io.h"
+
+int main() {
+  paddle_mobile::framework::Tensor input;
+  paddle_mobile::framework::Tensor output;
+  DLOG << 1;
+  SetupTensor<float>(&input, {1, 4, 60, 60}, static_cast<float>(0),
+                     static_cast<float>(1));
+  DLOG << 2;
+
+  auto out_ddim = paddle_mobile::framework::make_ddim({1, 4, 60, 60});
+  output.Resize(out_ddim);
+  DLOG << 3;
+  paddle_mobile::operators::sigmoid(&input, &output);
+  DLOG << 4;
+  auto *output_ptr = output.data<float>();
+  for (int j = 0; j < output.numel(); ++j) {
+    DLOG << " value of output: " << output_ptr[j];
+  }
+  DLOG << 5;
+  return 0;
+}