diff --git a/fpga/hls/load_calibration.cpp b/fpga/hls/load_calibration.cpp
index 22ab89f3..7d31963c 100644
--- a/fpga/hls/load_calibration.cpp
+++ b/fpga/hls/load_calibration.cpp
@@ -59,9 +59,11 @@ int load_calibration(ap_uint<256> *d_hbm_p0,
         for (int i = 0; i < 3 * (modules + storage_cells * modules); i++) {
                 if (in_mem_location[i] == 0)
                     return 2;
-                setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));
         }
 
+        for (int i = 0; i < 3 * (modules + storage_cells * modules); i++)
+                setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));
+
         for (int c = 0; c < 3; c++) {
             for (int m = 0; m < modules; m++) {
 #pragma HLS PIPELINE OFF
@@ -86,8 +88,10 @@ int load_calibration(ap_uint<256> *d_hbm_p0,
         for (int i = 0; i < 2 * modules; i++) {
             if (in_mem_location[i] == 0)
                 return 2;
-            setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));
         }
+            
+        for (int i = 0; i < 2 * modules; i++)
+            setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));
 
         for (int m = 0; m < modules; m++) {
 #pragma HLS PIPELINE OFF
@@ -106,9 +110,11 @@ int load_calibration(ap_uint<256> *d_hbm_p0,
         for (int i = 0; i < modules; i++) {
             if (in_mem_location[i] == 0)
                 return 2;
-            setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));
         }
 
+        for (int i = 0; i < modules; i++)        
+            setup_datamover(datamover_in_cmd, in_mem_location[i], RAW_MODULE_SIZE * sizeof(int16_t));        
+
         for (int m = 0; m < modules; m++) {
 #pragma HLS PIPELINE OFF
             size_t offset_hbm_0 = 20 * hbm_size_bytes / 32 + m * RAW_MODULE_SIZE * sizeof(int16_t) / 64;