improve the performance of gemm1*1s1_conv_add and gemm1*1s1_conv_add_bn_relu (#1754)

* add CPU_ARCH info, improve the performance of GEMM1*1s1

* improve the performance of gemm1*1s1_conv_add and gemm1*1s1_conv_add_bn_relu

* improve the performance of slidingwindow_bn_relu,slidingwindow_add,slidingwindow_add_bn_relu,gemm1*1s1_bn_relu,gemm1*1s1_add_relu

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

@@ -16,9 +16,12 @@ limitations under the License. */
 
 #include "operators/kernel/conv_add_bn_relu_kernel.h"
 #include <cmath>
+#include "framework/context.h"
 #include "operators/kernel/arm/convolution/conv_common.h"
 #include "operators/kernel/central-arm-func/conv_arm_func.h"
 #include "operators/math/element_wise.h"
+#include "operators/math/gemm/gemm1x1s1.h"
+#include "operators/math/slidingwindow_utils.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -64,6 +67,13 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
 
   // try to use faster depthwise conv
   switch (param->ExecMode()) {
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      use_slidingwindow_add_bn_relu = true;
+      break;
+    case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
+      use_gemm_add_bn_relu = true;
+      break;
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
       const std::vector<int> &paddings = param->Paddings();
@@ -84,7 +94,8 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
       break;
   }
 
-  if (could_use_faster_depthwise_conv_) {
+  if (could_use_faster_depthwise_conv_ || use_gemm_add_bn_relu ||
+      use_slidingwindow_add_bn_relu) {
     auto filter_data = param->Filter()->data<float>();
     auto filter_dim = param->Filter()->dims();
     int len = 1;
@@ -99,6 +110,16 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
             filter_data[i * step + k] * new_scale_ptr[i];
       }
     }
+    if (use_gemm_add_bn_relu) {
+      ARMArch arch = framework::CPUContext::Context()->get_arch();
+      math::gemm1x1s1_transform_weight(*param->Filter(), *param->Output(),
+                                       param->transformed_filter_,
+                                       param->groups, arch);
+    }
+    if (use_slidingwindow_add_bn_relu) {
+      math::slidingwindow_transform_weight<float>(*param->Filter(),
+                                                  param->transformed_filter_);
+    }
   }
 
   return true;
@@ -129,11 +150,15 @@ void ConvAddBNReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      fusion_has_been_computed = true;
+      GemmConv1x1s1<float, float>(param, param.NewBias()->data<float>(), true,
+                                  true);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, param.NewBias()->data<float>(),
+                                         true, true);
+      fusion_has_been_computed = true;
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",

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

@@ -30,6 +30,7 @@ bool ConvAddKernel<CPU, float>::Init(FusionConvAddParam<CPU> *param) {
 
 template <>
 void ConvAddKernel<CPU, float>::Compute(const FusionConvAddParam<CPU> &param) {
+  bool fusion_has_been_computed = false;
   switch (param.ExecMode()) {
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
@@ -45,22 +46,28 @@ void ConvAddKernel<CPU, float>::Compute(const FusionConvAddParam<CPU> &param) {
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      fusion_has_been_computed = true;
+      GemmConv1x1s1<float, float>(param, param.Bias()->data<float>(), true,
+                                  false);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, param.Bias()->data<float>(),
+                                         true, false);
+      fusion_has_been_computed = true;
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());
   }
-  if (param.Bias()->dims() == param.Output()->dims()) {
-    math::AddElememtWise<IDENTITY>(param.Output(), param.Bias(), param.Axis(),
-                                   param.Output());
-  } else {
-    math::AddChannelWise<IDENTITY>(param.Output(), param.Bias(),
-                                   param.Output());
+  if (!fusion_has_been_computed) {
+    if (param.Bias()->dims() == param.Output()->dims()) {
+      math::AddElememtWise<IDENTITY>(param.Output(), param.Bias(), param.Axis(),
+                                     param.Output());
+    } else {
+      math::AddChannelWise<IDENTITY>(param.Output(), param.Bias(),
+                                     param.Output());
+    }
   }
 }
 

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

@@ -31,6 +31,7 @@ bool ConvAddReluKernel<CPU, float>::Init(FusionConvAddReluParam<CPU> *param) {
 template <>
 void ConvAddReluKernel<CPU, float>::Compute(
     const FusionConvAddReluParam<CPU> &param) {
+  bool fusion_has_been_computed = false;
   switch (param.ExecMode()) {
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
@@ -46,21 +47,25 @@ void ConvAddReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      fusion_has_been_computed = true;
+      GemmConv1x1s1<float, float>(param, param.Bias()->data<float>(), true,
+                                  true);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, nullptr, false, false);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());
   }
-  if (param.Bias()->dims() == param.Output()->dims()) {
-    math::AddElememtWise<RELU>(param.Output(), param.Bias(), param.Axis(),
-                               param.Output());
-  } else {
-    math::AddChannelWise<RELU>(param.Output(), param.Bias(), param.Output());
+  if (!fusion_has_been_computed) {
+    if (param.Bias()->dims() == param.Output()->dims()) {
+      math::AddElememtWise<RELU>(param.Output(), param.Bias(), param.Axis(),
+                                 param.Output());
+    } else {
+      math::AddChannelWise<RELU>(param.Output(), param.Bias(), param.Output());
+    }
   }
 }
 

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

@@ -65,11 +65,11 @@ void ConvBNAddReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      GemmConv1x1s1<float, float>(param, nullptr, false, false);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, nullptr, false, false);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",

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

@@ -16,9 +16,12 @@ limitations under the License. */
 
 #include "operators/kernel/conv_bn_relu_kernel.h"
 #include <cmath>
+#include "framework/context.h"
 #include "operators/kernel/arm/convolution/conv_common.h"
 #include "operators/kernel/central-arm-func/conv_arm_func.h"
 #include "operators/math/element_wise.h"
+#include "operators/math/gemm/gemm1x1s1.h"
+#include "operators/math/slidingwindow_utils.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -57,12 +60,50 @@ bool ConvBNReluKernel<CPU, float>::Init(FusionConvBNReluParam<CPU> *param) {
   param->SetNewBias(new_bias);
 
   InitBaseConvKernel(param);
+
+  switch (param->ExecMode()) {
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
+      use_slidingwindow_bn_relu = true;
+      break;
+    case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
+      use_gemm_bn_relu = true;
+      break;
+  }
+
+  if (use_gemm_bn_relu || use_slidingwindow_bn_relu) {
+    auto filter_data = param->Filter()->data<float>();
+    auto filter_dim = param->Filter()->dims();
+    int len = 1;
+    for (int i = 0; i < filter_dim.size(); i++) {
+      len *= filter_dim[i];
+    }
+    int batch = filter_dim[0];
+    int step = len / batch;
+    for (int i = 0; i < batch; i++) {
+      for (int k = 0; k < step; k++) {
+        filter_data[i * step + k] =
+            filter_data[i * step + k] * new_scale_ptr[i];
+      }
+    }
+    if (use_gemm_bn_relu) {
+      ARMArch arch = framework::CPUContext::Context()->get_arch();
+      math::gemm1x1s1_transform_weight(*param->Filter(), *param->Output(),
+                                       param->transformed_filter_,
+                                       param->groups, arch);
+    }
+    if (use_slidingwindow_bn_relu) {
+      math::slidingwindow_transform_weight<float>(*param->Filter(),
+                                                  param->transformed_filter_);
+    }
+  }
   return true;
 }
 
 template <>
 void ConvBNReluKernel<CPU, float>::Compute(
     const FusionConvBNReluParam<CPU> &param) {
+  bool fusion_has_been_computed = false;
   switch (param.ExecMode()) {
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
@@ -78,18 +119,24 @@ void ConvBNReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      GemmConv1x1s1<float, float>(param, param.NewBias()->data<float>(), true,
+                                  true);
+      fusion_has_been_computed = true;
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, param.NewBias()->data<float>(),
+                                         true, true);
+      fusion_has_been_computed = true;
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());
   }
-  math::ScaleAddChannelWise<RELU>(param.Output(), param.NewScale(),
-                                  param.NewBias(), param.Output());
+  if (!fusion_has_been_computed) {
+    math::ScaleAddChannelWise<RELU>(param.Output(), param.NewScale(),
+                                    param.NewBias(), param.Output());
+  }
 }
 template class ConvBNReluKernel<CPU, float>;
 

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

@@ -55,11 +55,11 @@ void ConvKernel<CPU, float>::Compute(const ConvParam<CPU> &param) {
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      GemmConv1x1s1<float, float>(param, nullptr, false, false);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, nullptr, false, false);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",

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

@@ -46,11 +46,11 @@ void ConvReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      GemmConv1x1s1<float, float>(param, nullptr, false, false);
       break;
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_SLIDINGWINDOW3x3S2_FLOAT:
-      SlidingwindowConv3x3<float, float>(param);
+      SlidingwindowConv3x3<float, float>(param, nullptr, false, false);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",

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

@@ -77,7 +77,7 @@ void DWConvBNReluKernel<CPU, float>::Compute(
       GemmConv<float, float>(param);
       break;
     case ConvParam<CPU>::EXEC_GEMM1x1s1_FLOAT:
-      GemmConv1x1s1<float, float>(param);
+      GemmConv1x1s1<float, float>(param, nullptr, false, false);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",

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

@@ -140,7 +140,8 @@ void GemmConv(const ConvParam<CPU> &param) {
 }
 
 template <typename Itype, typename Otype>
-void GemmConv1x1s1(const ConvParam<CPU> &param) {
+void GemmConv1x1s1(const ConvParam<CPU> &param, const float *bias, bool is_bias,
+                   bool is_relu) {
   const Tensor *input = param.Input();
   Tensor filter = *param.transformed_filter_;
   Tensor *output = param.Output();
@@ -156,8 +157,6 @@ void GemmConv1x1s1(const ConvParam<CPU> &param) {
   const int hout = output->dims()[2];
   const int wout = output->dims()[3];
   const float *weights = filter.mutable_data<float>();
-  const float *bias = nullptr;
-
   int channel_size_out = wout * hout;
   int channel_size_in = win * hin;
   const int group = param.Groups();
@@ -165,8 +164,16 @@ void GemmConv1x1s1(const ConvParam<CPU> &param) {
   const int n = hout * wout;
   const int k = chin / group;
 
-  bool flag_relu = false;
-  bool flag_bias = false;
+  bool flag_relu = true;
+  bool flag_bias = true;
+
+  if (!is_bias) {
+    bias = nullptr;
+    flag_bias = false;
+  }
+  if (!is_relu) {
+    flag_relu = false;
+  }
   ARMArch arch = framework::CPUContext::Context()->get_arch();
   int hblock = math::get_hblock(arch);
 
@@ -322,7 +329,8 @@ void DepthwiseConv5x5(const ConvParam<CPU> &param) {
 }
 
 template <typename Itype, typename Otype>
-void SlidingwindowConv3x3(const ConvParam<CPU> &param) {
+void SlidingwindowConv3x3(const ConvParam<CPU> &param, const float *bias,
+                          bool is_bias, bool is_relu) {
   const Tensor *input = param.Input();
   const Tensor *filter = param.Filter();
   const std::vector<int> &paddings = param.Paddings();
@@ -334,23 +342,29 @@ void SlidingwindowConv3x3(const ConvParam<CPU> &param) {
     // math::SlidingwindowConv3x3s1<Itype, Otype>(input, filter, paddings,
     // output);
     math::SlidingwindowConv3x3s1Faster<Itype, Otype>(
-        input, param.transformed_filter_, paddings, output);
+        input, param.transformed_filter_, paddings, output, bias, is_bias,
+        is_relu);
   } else if (strides[0] == 2) {
     // math::SlidingwindowConv3x3s2<Itype, Otype>(input, filter, paddings,
     // output);
     math::SlidingwindowConv3x3s2Faster<Itype, Otype>(
-        input, param.transformed_filter_, paddings, output);
+        input, param.transformed_filter_, paddings, output, bias, is_bias,
+        is_relu);
   } else {
     GemmConv<Itype, Otype>(param);
   }
 }
 
 template void GemmConv<float, float>(const ConvParam<CPU> &param);
-template void GemmConv1x1s1<float, float>(const ConvParam<CPU> &param);
+template void GemmConv1x1s1<float, float>(const ConvParam<CPU> &param,
+                                          const float *bias, bool is_bias,
+                                          bool is_relu);
 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);
+template void SlidingwindowConv3x3<float, float>(const ConvParam<CPU> &param,
+                                                 const float *bias,
+                                                 bool is_bias, bool is_relu);
 
 template void GemmConv<int8_t, int32_t>(const ConvParam<CPU> &param);
 #ifndef __aarch64__

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

@@ -33,7 +33,8 @@ template <typename Itype, typename Otype>
 void GemmConv(const ConvParam<CPU> &param);
 
 template <typename Itype, typename Otype>
-void GemmConv1x1s1(const ConvParam<CPU> &param);
+void GemmConv1x1s1(const ConvParam<CPU> &param, const float *bias, bool is_bias,
+                   bool is_relu);
 
 template <int tile, int kernel>
 void WinogradConv3x3(const ConvParam<CPU> &param);
@@ -45,7 +46,8 @@ template <typename Itype, typename Otype>
 void DepthwiseConv5x5(const ConvParam<CPU> &param);
 
 template <typename Itype, typename Otype>
-void SlidingwindowConv3x3(const ConvParam<CPU> &param);
+void SlidingwindowConv3x3(const ConvParam<CPU> &param, const float *bias,
+                          bool is_bias, bool is_relu);
 
 void FasterDepthwiseConv3x3_bias_relu(const ConvParam<CPU> &param,
                                       const float *bias, bool flag_relu);

src/operators/kernel/conv_add_bn_relu_kernel.h

@@ -39,6 +39,8 @@ class ConvAddBNReluKernel
 
  private:
   bool could_use_faster_depthwise_conv_ = false;
+  bool use_gemm_add_bn_relu = false;
+  bool use_slidingwindow_add_bn_relu = false;
 };
 
 }  // namespace operators

src/operators/kernel/conv_bn_relu_kernel.h

@@ -36,6 +36,10 @@ class ConvBNReluKernel
  public:
   void Compute(const FusionConvBNReluParam<DeviceType> &param);
   bool Init(FusionConvBNReluParam<DeviceType> *param);
+
+ private:
+  bool use_gemm_bn_relu = false;
+  bool use_slidingwindow_bn_relu = false;
 };
 
 }  // namespace operators

src/operators/math/gemm/gemm1x1s1.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "operators/math/gemm/gemm1x1s1.h"
 #include <arm_neon.h>
 #include "framework/context.h"
-#include "iostream"
 
 namespace paddle_mobile {
 namespace operators {

src/operators/math/slidingwindow_conv3x3.cpp

@@ -3710,12 +3710,15 @@ void SlidingwindowConv3x3s2<float, float>(const framework::Tensor *input,
 template <>
 void SlidingwindowConv3x3s1Faster<float, float>(
     const framework::Tensor *input, framework::Tensor *filter,
-    const std::vector<int> &paddings, framework::Tensor *output) {
+    const std::vector<int> &paddings, framework::Tensor *output,
+    const float *bias, bool is_bias, bool is_relu) {
   const float *din = input->data<float>();
   float *dout = output->mutable_data<float>();
   const float *weights = filter->mutable_data<float>();
-  const float *bias = nullptr;
-  bool relu = false;
+  if (!is_bias) {
+    bias = nullptr;
+  }
+  bool relu = is_relu;
   const int num = input->dims()[0];
   const int chin = input->dims()[1];
   const int hin = input->dims()[2];
@@ -4623,12 +4626,15 @@ void SlidingwindowConv3x3s1Faster<float, float>(
 template <>
 void SlidingwindowConv3x3s2Faster<float, float>(
     const framework::Tensor *input, framework::Tensor *filter,
-    const std::vector<int> &paddings, framework::Tensor *output) {
+    const std::vector<int> &paddings, framework::Tensor *output,
+    const float *bias, bool is_bias, bool is_relu) {
   const float *din = input->data<float>();
   float *dout = output->mutable_data<float>();
   const float *weights = filter->mutable_data<float>();
-  const float *bias = nullptr;
-  bool relu = false;
+  if (!is_bias) {
+    bias = nullptr;
+  }
+  bool relu = is_relu;
   const int num = input->dims()[0];
   const int chin = input->dims()[1];
   const int hin = input->dims()[2];

src/operators/math/slidingwindow_conv3x3.h

@@ -37,13 +37,15 @@ template <typename Itype, typename Otype>
 void SlidingwindowConv3x3s1Faster(const framework::Tensor *input,
                                   framework::Tensor *filter,
                                   const std::vector<int> &paddings,
-                                  framework::Tensor *output);
+                                  framework::Tensor *output, const float *bias,
+                                  bool is_bias, bool is_relu);
 
 template <typename Itype, typename Otype>
 void SlidingwindowConv3x3s2Faster(const framework::Tensor *input,
                                   framework::Tensor *filter,
                                   const std::vector<int> &paddings,
-                                  framework::Tensor *output);
+                                  framework::Tensor *output, const float *bias,
+                                  bool is_bias, bool is_relu);
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile