diff --git a/src/blas_kernels.cu b/src/blas_kernels.cu
index 55d5e5d4..ed418c10 100644
--- a/src/blas_kernels.cu
+++ b/src/blas_kernels.cu
@@ -749,7 +749,9 @@ extern "C" void input_shortcut_gpu(float *in, int batch, int w1, int h1, int c1,
     if (sample < 1) sample = 1;
 
     int size = batch * minw * minh * minc;
-    input_shortcut_kernel << <cuda_gridsize(size), BLOCK, 0, get_cuda_stream() >> >(in, size, minw, minh, minc, stride, sample, batch, w1, h1, c1, add, w2, h2, c2, out);
+    //input_shortcut_kernel << <cuda_gridsize(size), BLOCK, 0, get_cuda_stream() >> >(in, size, minw, minh, minc, stride, sample, batch, w1, h1, c1, add, w2, h2, c2, out);
+    simple_copy_ongpu(w2 * h2 * c2 * batch, in, out);
+    shortcut_kernel << <cuda_gridsize(size), BLOCK, 0, get_cuda_stream() >> >(size, minw, minh, minc, stride, sample, batch, w1, h1, c1, add, w2, h2, c2, out);
     CHECK_CUDA(cudaPeekAtLastError());
 }