modify yolobox_cuda to support multiple runs (#2245)

* modify yolobox_cuda to support multiple runs test=develop

lite/backends/cuda/target_wrapper.cc

@@ -74,5 +74,16 @@ void TargetWrapperCuda::MemcpyAsync(void* dst,
   }
 }
 
+void TargetWrapperCuda::MemsetSync(void* devPtr, int value, size_t count) {
+  CUDA_CALL(cudaMemset(devPtr, value, count));
+}
+
+void TargetWrapperCuda::MemsetAsync(void* devPtr,
+                                    int value,
+                                    size_t count,
+                                    const stream_t& stream) {
+  CUDA_CALL(cudaMemsetAsync(devPtr, value, count, stream));
+}
+
 }  // namespace lite
 }  // namespace paddle

lite/backends/cuda/target_wrapper.h

@@ -59,6 +59,13 @@ class TargetWrapper<TARGET(kCUDA)> {
                           size_t size,
                           IoDirection dir,
                           const stream_t& stream);
+
+  static void MemsetSync(void* devPtr, int value, size_t count);
+
+  static void MemsetAsync(void* devPtr,
+                          int value,
+                          size_t count,
+                          const stream_t& stream);
 };
 }  // namespace lite
 }  // namespace paddle

lite/kernels/cuda/yolo_box_compute.cu

@@ -171,6 +171,10 @@ void YoloBoxCompute::Run() {
   const int* imgsize = ImgSize->data<int>();
   float* boxes = Boxes->mutable_data<float>(TARGET(kCUDA));
   float* scores = Scores->mutable_data<float>(TARGET(kCUDA));
+  TargetWrapperCuda::MemsetAsync(
+      boxes, 0, Boxes->numel() * sizeof(float), stream);
+  TargetWrapperCuda::MemsetAsync(
+      scores, 0, Scores->numel() * sizeof(float), stream);
 
   const int n = X->dims()[0];
   const int h = X->dims()[2];