Merge pull request #1031 from codeWorm2015/opencl

update cl image

src/framework/cl/cl_image.h

@@ -32,8 +32,13 @@ class CLImage {
 
   }
 
-  const DDim &TensorDim() {
-    return tensorDims_;
+  inline CLImage &Resize(const DDim &dims) {
+    tensorDims_ = dims;
+    return *this;
+  }
+
+  const DDim &dims() const {
+    return DDim();
   }
 
   std::vector<size_t> DefaultWorkSize() {

src/framework/op_registry.h

@@ -122,6 +122,9 @@ class OpRegistry {
 #define REGISTER_OPERATOR_FPGA(op_type, op_class) \
   REGISTER_OPERATOR(op_type, op_class, fpga, paddle_mobile::FPGA);
 
+#define REGISTER_OPERATOR_CL(op_type, op_class) \
+  REGISTER_OPERATOR(op_type, op_class, cl, paddle_mobile::GPU_CL);
+
 #define USE_OP(op_type, device_name)                                           \
   extern int TouchOpRegistrar_##op_type##_##device_name();                     \
   static int use_op_itself_##op_type##_##device_name __attribute__((unused)) = \
@@ -133,5 +136,7 @@ class OpRegistry {
 
 #define USE_OP_FPGA(op_type) USE_OP(op_type, fpga);
 
+#define USE_OP_CL(op_type) USE_OP(op_type, cl);
+
 }  // namespace framework
 }  // namespace paddle_mobile

src/operators/conv_op.cpp

@@ -62,4 +62,8 @@ REGISTER_OPERATOR_MALI_GPU(conv2d, ops::ConvOp);
 REGISTER_OPERATOR_FPGA(conv2d, ops::ConvOp);
 #endif
 
+#ifdef PADDLE_MOBILE_CL
+REGISTER_OPERATOR_CL(conv2d, ops::ConvOp);
+#endif
+
 #endif

src/operators/conv_op.h

@@ -45,11 +45,17 @@ class ConvOp : public framework::OperatorWithKernel<
 #ifdef PADDLE_MOBILE_CPU
 USE_OP_CPU(conv2d);
 #endif
+
 #ifdef PADDLE_MOBILE_MALI_GPU
 USE_OP_MALI_GPU(conv2d);
 #endif
+
 #ifdef PADDLE_MOBILE_FPGA
 USE_OP_FPGA(conv2d);
 #endif
 
+#ifdef PADDLE_MOBILE_CL
+USE_OP_CL(conv2d);
+#endif
+
 #endif