fix group like conv 3x3 (#3498)

* add test for imfix ,test=develop

* fix group like 3x3,test=develop

* add direct release build script, test=develop

* fix group like conv , test=develop

mobile/src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp

@@ -202,7 +202,6 @@ bool ConvAddBNReluKernel<GPU_CL, float>::Init(
     //      winograd_transform_weight<4, 3>(&this->cl_helper_, param->Filter());
     //
     //    } else {
-    param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
     param->Filter()->InitCLImage(cl_helper_.CLContext(),
                                  cl_helper_.CLCommandQueue());
     // std::cout << " input dim " << param->Input()->dims()[0] << "  "
@@ -218,7 +217,15 @@ bool ConvAddBNReluKernel<GPU_CL, float>::Init(
     //           param->Filter()->dims()[2]
     //           << " " << param->Filter()->dims()[3] << " " << std::endl;
 
-    this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file, build_options);
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file, build_options);
+    } else {
+      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
+                                 build_options);
+    }
     //    }
   } else {
     PADDLE_MOBILE_THROW_EXCEPTION(" not support ");
@@ -236,7 +243,7 @@ void ConvAddBNReluKernel<GPU_CL, float>::Compute(
                             param.NewScale(), param.NewBias());
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
-    // case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT:
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISEBASIC_FLOAT:
       ConvAddBnRelu(&this->cl_helper_, param, true, param.Bias(),

mobile/src/operators/kernel/cl/conv_add_kernel.cpp

@@ -96,10 +96,18 @@ bool ConvAddKernel<GPU_CL, float>::Init(FusionConvAddParam<GPU_CL> *param) {
     //
     //    } else {
 
-    param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
     param->Filter()->InitCLImage(cl_helper_.CLContext(),
                                  cl_helper_.CLCommandQueue());
-    this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file, build_options);
+
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file, build_options);
+    } else {
+      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
+                                 build_options);
+    }
     //    }
 
   } else if (param->Filter()->dims()[2] == 7 &&
@@ -130,6 +138,7 @@ void ConvAddKernel<GPU_CL, float>::Compute(
       WinogradConv3x3<4, 3>(&this->cl_helper_, param, false, param.Bias());
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW5x5_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISEBASIC_FLOAT:

mobile/src/operators/kernel/cl/conv_add_relu_kernel.cpp

@@ -96,7 +96,6 @@ bool ConvAddReluKernel<GPU_CL, float>::Init(
     //      winograd_transform_weight<4, 3>(&this->cl_helper_, param->Filter());
     //
     //    } else {
-    param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
     param->Filter()->InitCLImage(cl_helper_.CLContext(),
                                  cl_helper_.CLCommandQueue());
     // std::cout << " input dim " << param->Input()->dims()[0] << "  "
@@ -112,7 +111,16 @@ bool ConvAddReluKernel<GPU_CL, float>::Init(
     //           param->Filter()->dims()[2]
     //           << " " << param->Filter()->dims()[3] << " " << std::endl;
 
-    this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file, build_options);
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file, build_options);
+    } else {
+      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
+                                 build_options);
+    }
+
     //    }
 
   } else if (param->Filter()->dims()[2] == 7 &&
@@ -146,7 +154,7 @@ void ConvAddReluKernel<GPU_CL, float>::Compute(
       WinogradConv3x3<4, 3>(&this->cl_helper_, param, true, param.Bias());
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
-
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW5x5_FLOAT:
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW7x7_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:

mobile/src/operators/kernel/cl/conv_bn_relu_kernel.cpp

@@ -153,11 +153,18 @@ bool ConvBNReluKernel<GPU_CL, float>::Init(
     //      winograd_transform_weight<4, 3>(&this->cl_helper_, param->Filter());
     //
     //    } else {
-    param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
     param->Filter()->InitCLImage(cl_helper_.CLContext(),
                                  cl_helper_.CLCommandQueue());
 
-    this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file, build_options);
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file, build_options);
+    } else {
+      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
+                                 build_options);
+    }
     //    }
   } else {
     PADDLE_MOBILE_THROW_EXCEPTION(" not support ");
@@ -174,7 +181,7 @@ void ConvBNReluKernel<GPU_CL, float>::Compute(
                             param.NewScale(), param.NewBias());
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
-    // case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT:
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISEBASIC_FLOAT:
       ConvAddBnRelu(&this->cl_helper_, param, true, nullptr, param.NewScale(),

mobile/src/operators/kernel/cl/conv_kernel.cpp

@@ -90,7 +90,6 @@ bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
     //      winograd_transform_weight<4, 3>(&this->cl_helper_, param->Filter());
     //
     //    } else {
-    param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
     param->Filter()->InitCLImage(cl_helper_.CLContext(),
                                  cl_helper_.CLCommandQueue());
     // std::cout << " input dim " << param->Input()->dims()[0] << "  "
@@ -105,8 +104,15 @@ bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
     //           << param->Filter()->dims()[1] << " " <<
     //           param->Filter()->dims()[2]
     //           << " " << param->Filter()->dims()[3] << " " << std::endl;
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file);
+    } else {
+      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file);
+    }
 
-    this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file);
     //    }
     DLOG << "conv 3x3";
   } else if (param->Filter()->dims()[2] == 7 &&
@@ -132,7 +138,7 @@ void ConvKernel<GPU_CL, float>::Compute(const ConvParam<GPU_CL> &param) {
       WinogradConv3x3<4, 3>(&this->cl_helper_, param);
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
-    // case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT:
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW7x7_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISEBASIC_FLOAT:

mobile/src/operators/kernel/cl/conv_relu_kernel.cpp

@@ -96,29 +96,14 @@ bool ConvReluKernel<GPU_CL, float>::Init(FusionConvReluParam<GPU_CL> *param) {
     //      winograd_transform_weight<4, 3>(&this->cl_helper_, param->Filter());
     //
     //    } else {
-    if (param->Strides()[0] == 1 && param->Dilations()[0] == 1) {
-      param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3S1_FLOAT;
-      param->Filter()->InitCLImage(cl_helper_.CLContext(),
-                                   cl_helper_.CLCommandQueue());
-      this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
-                                 build_options);
+    param->Filter()->InitCLImage(cl_helper_.CLContext(),
+                                 cl_helper_.CLCommandQueue());
+    if (param->groups > 1) {
+      param->ExecMode() =
+          ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT;
+      this->cl_helper_.AddKernel("conv_3x3", conv_kernel_file, build_options);
     } else {
       param->ExecMode() = ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT;
-      param->Filter()->InitCLImage(cl_helper_.CLContext(),
-                                   cl_helper_.CLCommandQueue());
-      // std::cout << " input dim " << param->Input()->dims()[0] << "  "
-      //           << param->Input()->dims()[1] << "  "
-      //           << param->Input()->dims()[2] << "  "
-      //           << param->Input()->dims()[3] << "  " << std::endl;
-      // std::cout << " output dim " << param->Output()->dims()[0] << " "
-      //           << param->Output()->dims()[1] << " "
-      //           << param->Output()->dims()[2] << " "
-      //           << param->Output()->dims()[3] << " " << std::endl;
-      // std::cout << " filter dim " << param->Filter()->dims()[0] << " "
-      //           << param->Filter()->dims()[1] << " "
-      //           << param->Filter()->dims()[2] << " "
-      //           << param->Filter()->dims()[3] << " " << std::endl;
-
       this->cl_helper_.AddKernel("conv_3x3spl", conv_kernel_file,
                                  build_options);
     }
@@ -140,7 +125,7 @@ void ConvReluKernel<GPU_CL, float>::Compute(
       WinogradConv3x3<4, 3>(&this->cl_helper_, param, true);
       break;
     case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW1x1_FLOAT:
-    // case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_FLOAT:
+    case ConvParam<GPU_CL>::EXEC_SLIDINGWINDOW3x3_WITH_GROUP_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISE3x3_FLOAT:
     case ConvParam<GPU_CL>::EXEC_DEPTHWISEBASIC_FLOAT:
       ConvAddBnRelu(&this->cl_helper_, param, true);