!4770 [MS][LITE][Develop] fix codex for opencl depthwise

Merge pull request !4770 from wandongdong/master

mindspore/lite/src/runtime/kernel/opencl/cl/fp32/arithmetic_image2d.cl

+#define divide_no_check(a, b) (a/b)
 __constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
 
 __kernel void ElementAdd(__read_only image2d_t input_a, __read_only image2d_t input_b, __write_only image2d_t output,
@@ -49,10 +50,7 @@ __kernel void ElementDiv(__read_only image2d_t input_a, __read_only image2d_t in
 
   float4 a = read_imagef(input_a, smp_none, (int2)(X, Y));
   float4 b = read_imagef(input_b, smp_none, (int2)(X, Y));
-  if (b == 0) {
-    return;
-  }
-  write_imagef(output, (int2)(X, Y), a / b);
+  write_imagef(output, (int2)(X, Y), divide_no_check(a, b));
 }
 
 __kernel void BoardcastArith(__read_only image2d_t input_a, float weight, float bias, __write_only image2d_t output,

mindspore/lite/src/runtime/kernel/opencl/kernel/depthwise_conv2d.cc

@@ -102,7 +102,6 @@ int DepthwiseConv2dOpenCLKernel::InitBuffer() {
 
   allocator->UnmapBuffer(packed_weight_);
 
-  // init bias
   if (in_tensors_.size() == kInputSize2) {
     bias_data_ = reinterpret_cast<FLOAT_t *>(allocator->Malloc(C4NUM * CO4 * sizeof(FLOAT_t)));
     bias_data_ = reinterpret_cast<FLOAT_t *>(allocator->MapBuffer(bias_data_, CL_MAP_WRITE, nullptr, true));

mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.cc

@@ -15,7 +15,6 @@
  */
 
 #include "src/runtime/kernel/opencl/subgraph_opencl_kernel.h"
-#include <set>
 #include "src/runtime/opencl/opencl_executor.h"
 #include "src/runtime/opencl/opencl_runtime.h"
 #include "src/runtime/kernel/opencl/utils.h"
@@ -92,6 +91,8 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::tensor::Tensor *
     MS_ASSERT(parameter);
     if (parameter == nullptr) {
       MS_LOG(ERROR) << "SubGraphOpenCLKernel new parameter failed!";
+      delete new_tensor;
+      new_tensor = nullptr;
       return RET_ERROR;
     }
     parameter->src_format = src_format;
@@ -109,6 +110,10 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::tensor::Tensor *
     MS_ASSERT(in_convert_op);
     if (in_convert_op == nullptr) {
       MS_LOG(ERROR) << "SubGraphOpenCLKernel create op failed!";
+      delete new_tensor;
+      new_tensor = nullptr;
+      delete parameter;
+      parameter = nullptr;
       return RET_ERROR;
     }
     auto in_opencl_op = reinterpret_cast<OpenCLKernel *>(in_convert_op);
@@ -272,16 +277,16 @@ int SubGraphOpenCLKernel::UnInit() {
       delete tensor;
     }
   }
-  for (const auto parameter : in_parameters_) {
-    if (parameter != nullptr) {
-      delete parameter;
-    }
-  }
   for (const auto op : in_convert_ops_) {
     if (op != nullptr) {
       delete op;
     }
   }
+  for (const auto parameter : in_parameters_) {
+    if (parameter != nullptr) {
+      delete parameter;
+    }
+  }
   return RET_OK;
 }
 
@@ -290,18 +295,15 @@ int SubGraphOpenCLKernel::InferShape() { return RET_OK; }
 int SubGraphOpenCLKernel::ReSize() { return RET_OK; }
 
 int SubGraphOpenCLKernel::Run() {
+  auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   for (auto &tensor : in_tensors_) {
     allocator_->UnmapBuffer(tensor->Data());
   }
 
   lite::opencl::OpenCLExecutor executor;
   executor.Run(in_tensors_, out_tensors_, nodes_, allocator_);
+  ocl_runtime->SyncCommandQueue();
 
-  for (auto &tensor : out_tensors_) {
-    void *data = allocator_->MapBuffer(tensor->Data(), CL_MAP_READ, nullptr, true);
-    tensor->SetData(data);
-  }
   return RET_OK;
 }
-
 }  // namespace mindspore::kernel