Merge branch 'develop' into my-cool-stuff

src/operators/kernel/arm/convolution/conv_add_bn_relu_kernel.cpp

@@ -78,6 +78,10 @@ void ConvAddBNReluKernel<CPU, float>::Compute(
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/arm/convolution/conv_add_kernel.cpp

@@ -32,10 +32,8 @@ template <>
 void ConvAddKernel<CPU, float>::Compute(const FusionConvAddParam<CPU> &param) {
   switch (param.ExecMode()) {
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
-      break;
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
-      math::DepthwiseConv3x3S2<float, float>(*param.Input(), *param.Filter(),
-                                             param.Paddings(), param.Output());
+      DepthwiseConv3x3<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_DEPTHWISE5x5_FLOAT:
       DepthwiseConv5x5<float, float>(param);
@@ -46,6 +44,10 @@ void ConvAddKernel<CPU, float>::Compute(const FusionConvAddParam<CPU> &param) {
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/arm/convolution/conv_add_relu_kernel.cpp

@@ -45,6 +45,10 @@ void ConvAddReluKernel<CPU, float>::Compute(
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/arm/convolution/conv_bn_add_relu_kernel.cpp

@@ -76,6 +76,10 @@ void ConvBNAddReluKernel<CPU, float>::Compute(
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/arm/convolution/conv_bn_relu_kernel.cpp

@@ -75,6 +75,10 @@ void ConvBNReluKernel<CPU, float>::Compute(
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/arm/convolution/conv_common.cpp

@@ -57,8 +57,8 @@ void InitBaseConvKernel(ConvParam<CPU> *param) {
                param->Dilations()[0] == param->Dilations()[1] &&
                param->Strides()[0] == 1 && param->Dilations()[0] == 1
 #if 1
-               && (param->Input()->dims()[1] >= 4 ||
-                   param->Output()->dims()[1] >= 16)
+               && (param->Input()->dims()[1] >= 8 &&
+                   param->Output()->dims()[1] >= 8)
 #endif
     ) {
       param->ExecMode() = ConvParam<CPU>::EXEC_WINOGRAD3X3_FLOAT;
@@ -66,6 +66,26 @@ void InitBaseConvKernel(ConvParam<CPU> *param) {
       param->transformed_filter_ = new framework::LoDTensor;
       operators::math::winograd_transform_weight<8, 3>(
           *param->Filter(), param->transformed_filter_);
+    } else if (conv3x3 && !depth3x3 &&
+               param->Strides()[0] == param->Strides()[1] &&
+               param->Dilations()[0] == param->Dilations()[1] &&
+               param->Strides()[0] == 1 && param->Dilations()[0] == 1
+#if 1
+               && (param->Input()->dims()[2] >= 48 &&
+                   param->Output()->dims()[1] <= 24)
+#endif
+    ) {
+      param->ExecMode() = ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT;
+    } else if (conv3x3 && !depth3x3 &&
+               param->Strides()[0] == param->Strides()[1] &&
+               param->Dilations()[0] == param->Dilations()[1] &&
+               param->Strides()[0] == 2 && param->Dilations()[0] == 1
+#if 1
+               && (param->Input()->dims()[2] >= 48 &&
+                   param->Output()->dims()[1] <= 24)
+#endif
+    ) {
+      param->ExecMode() = ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT;
     } else {
       param->ExecMode() = ConvParam<CPU>::EXEC_GEMM_FLOAT;
     }

src/operators/kernel/arm/convolution/conv_kernel.cpp

@@ -54,6 +54,10 @@ void ConvKernel<CPU, float>::Compute(const ConvParam<CPU> &param) {
     case ConvParam<CPU>::EXEC_GEMM_FLOAT:
       GemmConv<float, float>(param);
       break;
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      SlidingwindowConv3x3<float, float>(param);
+      break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());

src/operators/kernel/central-arm-func/conv_arm_func.cpp

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "operators/math/im2col.h"
 #include "operators/math/math_function.h"
 #include "operators/math/pad.h"
+#include "operators/math/slidingwindow_conv3x3.h"
 #include "operators/math/vol2col.h"
 #include "operators/math/winograd/winograd_transform.h"
 #include "operators/op_param.h"
@@ -232,10 +233,29 @@ void DepthwiseConv5x5(const ConvParam<CPU> &param) {
   }
 }
 
+template <typename Itype, typename Otype>
+void SlidingwindowConv3x3(const ConvParam<CPU> &param) {
+  const Tensor *input = param.Input();
+  const Tensor *filter = param.Filter();
+  const std::vector<int> &paddings = param.Paddings();
+  const std::vector<int> &strides = param.Strides();
+  Tensor *output = param.Output();
+  output->mutable_data<Otype>();
+
+  if (strides[0] == 1) {
+    math::SlidingwindowConv3x3s1<Itype, Otype>(input, filter, paddings, output);
+  } else if (strides[0] == 2) {
+    math::SlidingwindowConv3x3s2<Itype, Otype>(input, filter, paddings, output);
+  } else {
+    GemmConv<Itype, Otype>(param);
+  }
+}
+
 template void GemmConv<float, float>(const ConvParam<CPU> &param);
 template void WinogradConv3x3<8, 3>(const ConvParam<CPU> &param);
 template void DepthwiseConv3x3<float, float>(const ConvParam<CPU> &param);
 template void DepthwiseConv5x5<float, float>(const ConvParam<CPU> &param);
+template void SlidingwindowConv3x3<float, float>(const ConvParam<CPU> &param);
 
 #ifndef __aarch64__
 template void GemmConv<int8_t, int32_t>(const ConvParam<CPU> &param);

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

@@ -41,6 +41,9 @@ void DepthwiseConv3x3(const ConvParam<CPU> &param);
 template <typename Itype, typename Otype>
 void DepthwiseConv5x5(const ConvParam<CPU> &param);
 
+template <typename Itype, typename Otype>
+void SlidingwindowConv3x3(const ConvParam<CPU> &param);
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/math/slidingwindow_conv3x3.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 <algorithm>
+#include <vector>
+#include "framework/tensor.h"
+
+namespace paddle_mobile {
+namespace operators {
+namespace math {
+template <typename Itype, typename Otype>
+void SlidingwindowConv3x3s1(const framework::Tensor *input,
+                            const framework::Tensor *filter,
+                            const std::vector<int> &paddings,
+                            framework::Tensor *output);
+
+template <typename Itype, typename Otype>
+void SlidingwindowConv3x3s2(const framework::Tensor *input,
+                            const framework::Tensor *filter,
+                            const std::vector<int> &paddings,
+                            framework::Tensor *output);
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/op_param.h

@@ -476,6 +476,8 @@ class ConvParam : public OpParam {
     EXEC_GEMM_INT8,
     EXEC_DEPTHWISE3x3_INT8,
     EXEC_DEPTHWISE5x5_INT8,
+    EXEC_SLIDINGWINDOW3x3S1_FLOAT,
+    EXEC_SLIDINGWINDOW3x3S2_FLOAT,
   };
 
   ExecMode &ExecMode() const { return exec_mode_; }