Avoid using dynamic array in cuda kernel

paddle/fluid/operators/argsort_op.cc

@@ -38,8 +38,8 @@ class ArgsortOp : public framework::OperatorWithKernel {
                    "dimension %d.",
                    axis, num_dims);
     PADDLE_ENFORCE(in_dims.size() + axis >= 0,
-                   "Attr(axis) %d of ArgsortOp plus the number of Input(X)'s "
-                   "dimensions %d must be nonnegative.",
+                   "Attr(axis) %d of ArgsortOp plus the rank %d of Input(X) "
+                   "must be nonnegative.",
                    axis, in_dims.size());
 
     ctx->SetOutputDim("Out", in_dims);

paddle/fluid/operators/argsort_op.cu

@@ -31,8 +31,9 @@ __global__ void ComputeTargetIdx(const int64_t* in_dims, int dims_size,
                                  int64_t* med_ids) {
   int64_t index = threadIdx.x + blockDim.x * blockIdx.x;
   if (index < n) {
-    int64_t* shape_out_axis = new int64_t[dims_size - 1];
-    int64_t* dims_out_axis = new int64_t[dims_size - 1];
+    const int max_rank = 9;  // Max rank of a tensor allow in Fluid
+    int64_t shape_out_axis[max_rank - 1] = {0};
+    int64_t dims_out_axis[max_rank - 1] = {0};
     int64_t tmp = index;
     int64_t pos_in_axis = 0;
     int64_t i = dims_size - 2;
@@ -57,8 +58,6 @@ __global__ void ComputeTargetIdx(const int64_t* in_dims, int dims_size,
     int64_t traget_idx = group * dim_axis + pos_in_axis;
     trg_idx[index] = traget_idx;
     med_ids[traget_idx] = pos_in_axis;
-    delete[] shape_out_axis;
-    delete[] dims_out_axis;
   }
 }