Fix run gpu model in adreno 630

mindspore/lite/src/litert/kernel/gpu/opencl/opencl_allocator.cc

@@ -213,7 +213,11 @@ void *OpenCLAllocator::_Malloc(MemType mem_type, void *data, size_t size, const
 
   total_size_ += size;
   const uint64_t max_size = ocl_runtime_->GetGlobalMemSize() * 0.8;
-  UNLOCK_AND_RETURN_NULL(total_size_ >= max_size, nullptr);
+  if (total_size_ >= max_size) {
+    is_oversize_ = true;
+  } else {
+    is_oversize_ = false;
+  }
 
   cl::Buffer *buffer = nullptr;
   cl::Image2D *image = nullptr;

mindspore/lite/src/litert/kernel/gpu/opencl/opencl_allocator.h

@@ -83,6 +83,7 @@ class OpenCLAllocator : public mindspore::Allocator {
     }
     return ptr;
   }
+  bool IsOverSize() { return is_oversize_; }
 
  private:
   void Lock();
@@ -117,6 +118,7 @@ class OpenCLAllocator : public mindspore::Allocator {
   // 6 is empirical value
   int shift_factor_ = 6;
   bool lock_flag_ = true;
+  bool is_oversize_{false};
 };
 }  // namespace mindspore::lite::opencl
 

mindspore/lite/src/litert/kernel/opencl/opencl_kernel.cc

@@ -277,6 +277,7 @@ int OpenCLKernel::PreProcess() {
     }
     output->ResetRefCount();
   }
+  is_oversize_kernel_ = ocl_runtime_->GetAllocator()->IsOverSize();
   return RET_OK;
 }
 

mindspore/lite/src/litert/kernel/opencl/opencl_kernel.h

@@ -284,7 +284,12 @@ class OpenCLKernel : public LiteKernel {
   int PreProcess() override;
   int ReSize() override;
   int Run() override { return RET_ERROR; }
-  int PostProcess() override { return RET_OK; }
+  int PostProcess() override {
+    if (is_oversize_kernel_) {
+      return FreeInWorkTensor();
+    }
+    return RET_OK;
+  }
 
   bool MallocDataDone();
   std::string OpenCLKernelHeader();
@@ -336,6 +341,7 @@ class OpenCLKernel : public LiteKernel {
   cl::Event event_;
   void *restore_quant_data_{nullptr};
   bool dequant_flag_{false};
+  bool is_oversize_kernel_{false};
 
  private:
   lite::opencl::OpenCLRuntimeInnerWrapper ocl_runtime_wrap_;