fix(dnn): fix the sync problem in some kernels

GitOrigin-RevId: df3f7dc51b

dnn/src/cuda/convolution3d/backward_filter/inplace_matmul_impl.cu

@@ -355,7 +355,8 @@ void convolution3d::exec_inplace_matmul_bwd_filter(
     } else {
         BX = BY = 16;
     }
-    cudaMemset(grad, 0, OC * IC * FD * FH * FW * sizeof(float));
+    cuda_check(
+            cudaMemsetAsync(grad, 0, OC * IC * FD * FH * FW * sizeof(float), stream));
     dim3 blocks(DIVUP(n, 4 * BX), DIVUP(m, 4 * BY), N);
     dim3 threads(BX, BY);
 #define DISPATCH_BX_BY(BX, BY)                                                      \

dnn/src/cuda/padding/padding.cu

@@ -224,7 +224,7 @@ void padding_backward_proxy(
         params.offsets[i * 2 + 1] = offsets[i * 2 + 1];
     }
 
-    cudaMemset(dst.raw_ptr(), 0, dst.layout.access_bytes());
+    cuda_check(cudaMemsetAsync(dst.raw_ptr(), 0, dst.layout.access_bytes(), stream));
 
     void (*bwd_kern)(const size_t, const size_t, const T* const, T* const, ShapeParams);