[OPENCL] remove conv redundant's for opencl kernel. test=develop (#3924)

remove conv redundant's for opencl kernel.

lite/backends/opencl/cl_context.cc

@@ -119,7 +119,7 @@ cl::NDRange CLContext::DefaultWorkSize(const CLImage &image) {
   }
 }
 
-cl::NDRange CLContext::LocalWorkSizeTurn(cl::NDRange global_work_size,
+cl::NDRange CLContext::LocalWorkSizeTune(cl::NDRange global_work_size,
                                          size_t max_work_size,
                                          int divisor) {
   int preferred_lws = 0;
@@ -157,7 +157,7 @@ cl::NDRange CLContext::LocalWorkSizeTurn(cl::NDRange global_work_size,
                      static_cast<size_t>(gws0)};
 #endif
 }
-cl::NDRange CLContext::LocalWorkSizeTurnReverse(cl::NDRange global_work_size,
+cl::NDRange CLContext::LocalWorkSizeTuneReverse(cl::NDRange global_work_size,
                                                 size_t max_work_size,
                                                 int divisor) {
   int preferred_lws = 0;

lite/backends/opencl/cl_context.h

@@ -62,10 +62,10 @@ class CLContext {
 
   cl::NDRange LocalWorkSize(cl::NDRange global_work_size, size_t max_work_size);
 
-  cl::NDRange LocalWorkSizeTurn(cl::NDRange global_work_size,
+  cl::NDRange LocalWorkSizeTune(cl::NDRange global_work_size,
                                 size_t max_work_size,
                                 int divitor = 2);
-  cl::NDRange LocalWorkSizeTurnReverse(cl::NDRange global_work_size,
+  cl::NDRange LocalWorkSizeTuneReverse(cl::NDRange global_work_size,
                                        size_t max_work_size,
                                        int divitor = 2);
   bool IsArmMali();

lite/backends/opencl/cl_kernel/image/conv2d_1x1_opt_kernel.cl

@@ -6,9 +6,7 @@ __kernel void conv2d_1x1_opt(
     __private const int global_size_dim2,
     __read_only image2d_t input_image,
     __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
     __read_only image2d_t bias,
-#endif
 #ifdef BATCH_NORM
     __read_only image2d_t new_scale,
     __read_only image2d_t new_biase,
@@ -284,9 +282,7 @@ __kernel void conv2d_1x1_simple(
     __private const int global_size_dim2,
     __read_only image2d_t input_image,
     __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
     __read_only image2d_t bias,
-#endif
 #ifdef BATCH_NORM
     __read_only image2d_t new_scale,
     __read_only image2d_t new_biase,

lite/backends/opencl/cl_kernel/image/conv2d_3x3_kernel.cl

@@ -19,9 +19,7 @@ __kernel void conv2d_3x3(__private const int global_size_dim0,
                          __private const int global_size_dim2,
                          __read_only image2d_t input_image,
                          __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                          __read_only image2d_t bias,
-#endif
                          __write_only image2d_t output_image,
                          __private const int stride,
                          __private const int offset,

lite/backends/opencl/cl_kernel/image/conv2d_3x3_opt_kernel.cl

@@ -19,9 +19,7 @@ __kernel void conv2d_3x3_opt(__private const int item_ch,
                              __private const int item_h,
                              __read_only image2d_t input_image,
                              __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                              __read_only image2d_t bias,
-#endif
                              __write_only image2d_t output_image,
                              __private const int stride,
                              __private const int pad,
@@ -264,9 +262,7 @@ __kernel void conv2d_3x3_multi_batch(__private const int item_ch,
                                      __private const int item_h,
                                      __read_only image2d_t input_image,
                                      __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                                      __read_only image2d_t bias,
-#endif
                                      __write_only image2d_t output_image,
                                      __private const int stride,
                                      __private const int pad,

lite/backends/opencl/cl_kernel/image/conv2d_5x5_kernel.cl

@@ -5,9 +5,7 @@ __kernel void conv2d_5x5(__private const int global_size_dim0,
                          __private const int global_size_dim2,
                          __read_only image2d_t input_image,
                          __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                          __read_only image2d_t bias,
-#endif
 #ifdef BATCH_NORM
                          __read_only image2d_t new_scale,
                          __read_only image2d_t new_biase,

lite/backends/opencl/cl_kernel/image/conv2d_5x5_opt_kernel.cl

@@ -20,9 +20,7 @@ __kernel void conv2d_5x5_opt(__private const int item_ch,
                              __private const int item_h,
                              __read_only image2d_t input_image,
                              __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                              __read_only image2d_t bias,
-#endif
                              __write_only image2d_t output_image,
                              __private const int stride,
                              __private const int pad,
@@ -268,9 +266,7 @@ __kernel void conv2d_5x5_multi_batch(__private const int item_ch,
                                      __private const int item_h,
                                      __read_only image2d_t input_image,
                                      __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                                      __read_only image2d_t bias,
-#endif
                                      __write_only image2d_t output_image,
                                      __private const int stride,
                                      __private const int pad,
@@ -513,4 +509,4 @@ __kernel void conv2d_5x5_multi_batch(__private const int item_ch,
                    (int2)(out_w_base_id + out_w_id4, item_h_id),
                    output[4]);
   }
-}
\ No newline at end of file
+}

lite/backends/opencl/cl_kernel/image/conv2d_7x7_kernel.cl

@@ -5,9 +5,7 @@ __kernel void conv2d_7x7(__private const int global_size_dim0,
                          __private const int global_size_dim2,
                          __read_only image2d_t input_image,
                          __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                          __read_only image2d_t bias,
-#endif
 #ifdef BATCH_NORM
                          __read_only image2d_t new_scale,
                          __read_only image2d_t new_biase,

lite/backends/opencl/cl_kernel/image/conv2d_7x7_opt_kernel.cl

@@ -20,9 +20,7 @@ __kernel void conv2d_7x7_opt(__private const int item_ch,
                              __private const int item_h,
                              __read_only image2d_t input_image,
                              __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                              __read_only image2d_t bias,
-#endif
                              __write_only image2d_t output_image,
                              __private const int stride,
                              __private const int pad,
@@ -268,9 +266,7 @@ __kernel void conv2d_7x7_multi_batch(__private const int item_ch,
                                      __private const int item_h,
                                      __read_only image2d_t input_image,
                                      __read_only image2d_t filter_image,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                                      __read_only image2d_t bias,
-#endif
                                      __write_only image2d_t output_image,
                                      __private const int stride,
                                      __private const int pad,
@@ -513,4 +509,4 @@ __kernel void conv2d_7x7_multi_batch(__private const int item_ch,
                    (int2)(out_w_base_id + out_w_id4, item_h_id),
                    output[4]);
   }
-}
\ No newline at end of file
+}

lite/backends/opencl/cl_kernel/image/depthwise_conv2d_basic_kernel.cl

@@ -19,9 +19,7 @@ __kernel void depth_conv2d(__private const int global_size_dim0,
                            __private const int global_size_dim2,
                            __read_only image2d_t input,
                            __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                            __read_only image2d_t bias,
-#endif
 #ifdef BATCH_NORM
                            __read_only image2d_t new_scale,
                            __read_only image2d_t new_biase,

lite/backends/opencl/cl_kernel/image/depthwise_conv2d_kernel.cl

@@ -20,9 +20,7 @@ __kernel void depth_conv2d_3x3(
     __private const int global_size_dim2,
     __read_only image2d_t input,
     __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
     __read_only image2d_t bias,
-#endif
     __write_only image2d_t output_image,
     __private const int stride,
     __private const int offset,
@@ -249,9 +247,7 @@ __kernel void depth_conv2d_3x3s1(__private const int ou_ch_blk,
                                  __private const int ou_nh,
                                  __read_only image2d_t input,
                                  __read_only image2d_t filter,
-#if defined(BIASE_CH) || defined(BIASE_ELE)
                                  __read_only image2d_t bias,
-#endif
                                  __write_only image2d_t output_image,
                                  __private const int stride,
                                  __private const int pad,

lite/kernels/opencl/conv_image_compute.h

@@ -33,6 +33,7 @@ namespace paddle {
 namespace lite {
 namespace kernels {
 namespace opencl {
+
 class ConvImageCompute : public KernelLite<TARGET(kOpenCL),
                                            PRECISION(kFP16),
                                            DATALAYOUT(kImageDefault)> {
@@ -42,8 +43,11 @@ class ConvImageCompute : public KernelLite<TARGET(kOpenCL),
 
   void PrepareForRun() override;
 
+  void ReInitWhenNeeded() override;
+
   void Run() override;
-  double Turn(int times = 5);
+
+  double Tune(int times = 5);
 
 #ifdef LITE_WITH_PROFILE
   void SetProfileRuntimeKernelInfo(paddle::lite::profile::OpCharacter* ch) {
@@ -56,16 +60,20 @@ class ConvImageCompute : public KernelLite<TARGET(kOpenCL),
 #endif
 
  private:
-  void Conv2d1x1opt(bool is_turn = false);
-  void Conv2d3x3(bool is_turn = false);
-  void Conv2d3x3opt(bool is_turn = false);
-  void Conv2d5x5(bool is_turn = false);
-  void Conv2d5x5opt(bool is_turn = false);
-  void Conv2d7x7(bool is_turn = false);
-  void Conv2d7x7opt(bool is_turn = false);
-  void DepthwiseConv2d3x3s1(bool is_turn = false);
-  void DepthwiseConv2d3x3(bool is_turn = false);
-  void DepthwiseConv2d(bool is_turn = false);
+  void PrintConvInfo();
+  void GetGlobalWorkSize();
+  void Conv2d1x1opt(bool enable_tune = false);
+  void Conv2d3x3(bool enable_tune = false);
+  void Conv2d3x3opt(bool enable_tune = false);
+  void Conv2d5x5(bool enable_tune = false);
+  void Conv2d5x5opt(bool enable_tune = false);
+  void Conv2d7x7(bool enable_tune = false);
+  void Conv2d7x7opt(bool enable_tune = false);
+  void DepthwiseConv2d3x3s1(bool enable_tune = false);
+  void DepthwiseConv2d3x3(bool enable_tune = false);
+  void DepthwiseConv2d(bool enable_tune = false);
+
+  param_t* conv_param_{nullptr};
 
   kernel_t impl_;
   std::vector<std::string> kernel_func_names_{};
@@ -79,19 +87,72 @@ class ConvImageCompute : public KernelLite<TARGET(kOpenCL),
   std::unique_ptr<Tensor> tensor_hold_bias_image_{nullptr};
   cl::NDRange global_work_size_ = cl::NDRange{
       static_cast<size_t>(1), static_cast<size_t>(1), static_cast<size_t>(1)};
+
+  // opencl kernel args
   int c_blk_ = 1;
   int w_blk_ = 1;
   int nh_blk_ = 1;
 
+  const cl::Image2D* input_image_p_{nullptr};
+  const cl::Image2D* filter_image_p_{nullptr};
+  const cl::Image2D* bias_image_p_{nullptr};
+  const cl::Image2D* output_image_p_{nullptr};
+
+  int stride_h_{-1};
+  int stride_w_{-1};
+
+  int dilation_h_{-1};
+  int dilation_w_{-1};
+
+  int pad_up_{-1};
+  int pad_down_{-1};
+  int pad_left_{-1};
+  int pad_right_{-1};
+
+  int offset_{-1};
+  int groups_{-1};
+  bool relu_fused_{false};
+  bool has_bias_{false};
+
+  int input_tensor_n_{-1};
+  int input_tensor_c_{-1};
+  int input_tensor_h_{-1};
+  int input_tensor_w_{-1};
+  int input_image_h_{-1};
+  int input_image_w_{-1};
+  int input_c_block_{-1};
+
+  int output_tensor_n_{-1};
+  int output_tensor_c_{-1};
+  int output_tensor_h_{-1};
+  int output_tensor_w_{-1};
+  int output_image_h_{-1};
+  int output_image_w_{-1};
+
+  int filter_tensor_n_{-1};
+  int filter_tensor_c_{-1};
+  int filter_tensor_h_{-1};
+  int filter_tensor_w_{-1};
+  int filter_image_h_{-1};
+  int filter_image_w_{-1};
+
+  int bias_image_h_{-1};
+  int bias_image_w_{-1};
+
   int default_c_blk_ = 1;
   int default_w_blk_ = 1;
   int default_nh_blk_ = 1;
+  // =================
+
+  DDim last_input_dims_{};
+  bool is_first_epoch_for_run_{true};
 
   cl::Kernel kernel_;
+  cl_int status_;
   cl::NDRange local_work_size_ = cl::NDRange{
       static_cast<size_t>(1), static_cast<size_t>(1), static_cast<size_t>(1)};
   bool use_lws_{true};
-  bool use_turn_{false};
+  bool use_tune_{false};
 };
 
 }  // namespace opencl