diff --git a/openfpga/src/base/openfpga_bitstream.cpp b/openfpga/src/base/openfpga_bitstream.cpp
index ea390ae08..9dff1c7a8 100644
--- a/openfpga/src/base/openfpga_bitstream.cpp
+++ b/openfpga/src/base/openfpga_bitstream.cpp
@@ -122,6 +122,7 @@ int write_fabric_bitstream(const OpenfpgaContext& openfpga_ctx,
     /* By default, output in plain text format */
     status = write_fabric_bitstream_to_text_file(openfpga_ctx.bitstream_manager(),
                                                  openfpga_ctx.fabric_bitstream(),
+                                                 openfpga_ctx.blwl_shift_register_banks(),
                                                  openfpga_ctx.arch().config_protocol,
                                                  openfpga_ctx.fabric_global_port_info(),
                                                  cmd_context.option_value(cmd, opt_file),
diff --git a/openfpga/src/base/openfpga_context.h b/openfpga/src/base/openfpga_context.h
index 3719a2fa7..f4d356550 100644
--- a/openfpga/src/base/openfpga_context.h
+++ b/openfpga/src/base/openfpga_context.h
@@ -66,7 +66,7 @@ class OpenfpgaContext : public Context  {
     const openfpga::DeviceRRGSB& device_rr_gsb() const { return device_rr_gsb_; }
     const openfpga::MuxLibrary& mux_lib() const { return mux_lib_; }
     const openfpga::DecoderLibrary& decoder_lib() const { return decoder_lib_; }
-    const openfpga::MemoryBankShiftRegisterBanks& blwl_shift_register_banks() { return blwl_sr_banks_; }
+    const openfpga::MemoryBankShiftRegisterBanks& blwl_shift_register_banks() const { return blwl_sr_banks_; }
     const openfpga::TileDirect& tile_direct() const { return tile_direct_; }
     const openfpga::ModuleManager& module_graph() const { return module_graph_; }
     const openfpga::FlowManager& flow_manager() const { return flow_manager_; }
diff --git a/openfpga/src/base/openfpga_verilog.cpp b/openfpga/src/base/openfpga_verilog.cpp
index 2f3dd892f..4d4a615e7 100644
--- a/openfpga/src/base/openfpga_verilog.cpp
+++ b/openfpga/src/base/openfpga_verilog.cpp
@@ -107,6 +107,7 @@ int write_full_testbench(const OpenfpgaContext& openfpga_ctx,
   return fpga_verilog_full_testbench(openfpga_ctx.module_graph(),
                                      openfpga_ctx.bitstream_manager(),
                                      openfpga_ctx.fabric_bitstream(),
+                                     openfpga_ctx.blwl_shift_register_banks(),
                                      g_vpr_ctx.atom(),
                                      g_vpr_ctx.placement(),
                                      pin_constraints,
diff --git a/openfpga/src/fabric/memory_bank_shift_register_banks.cpp b/openfpga/src/fabric/memory_bank_shift_register_banks.cpp
index 0a506b9ae..39eaa8ad3 100644
--- a/openfpga/src/fabric/memory_bank_shift_register_banks.cpp
+++ b/openfpga/src/fabric/memory_bank_shift_register_banks.cpp
@@ -6,6 +6,10 @@
 /* begin namespace openfpga */
 namespace openfpga {
 
+ModuleManager::region_range MemoryBankShiftRegisterBanks::regions() const {
+  return vtr::make_range(config_region_ids_.begin(), config_region_ids_.end());
+}
+
 std::vector<size_t> MemoryBankShiftRegisterBanks::bl_bank_unique_sizes() const {
   std::vector<size_t> unique_sizes;
   for (const auto& region : bl_bank_data_ports_) {
@@ -203,6 +207,10 @@ BasicPort MemoryBankShiftRegisterBanks::find_wl_shift_register_bank_data_port(co
 }
 
 void MemoryBankShiftRegisterBanks::resize_regions(const size_t& num_regions) {
+  config_region_ids_.resize(num_regions);
+  for (size_t iregion = 0; iregion < num_regions; ++iregion) {
+    config_region_ids_[ConfigRegionId(iregion)] = ConfigRegionId(iregion);
+  }
   bl_bank_ids_.resize(num_regions);
   bl_bank_data_ports_.resize(num_regions);
   bl_bank_modules_.resize(num_regions);
diff --git a/openfpga/src/fabric/memory_bank_shift_register_banks.h b/openfpga/src/fabric/memory_bank_shift_register_banks.h
index fd226ded0..32df3444e 100644
--- a/openfpga/src/fabric/memory_bank_shift_register_banks.h
+++ b/openfpga/src/fabric/memory_bank_shift_register_banks.h
@@ -23,6 +23,7 @@ namespace openfpga {
  ******************************************************************************/
 class MemoryBankShiftRegisterBanks {
   public: /* Accessors: aggregates */
+    ModuleManager::region_range regions() const;
     FabricKey::fabric_bit_line_bank_range bl_banks(const ConfigRegionId& region_id) const;
     FabricKey::fabric_word_line_bank_range wl_banks(const ConfigRegionId& region_id) const;
   public: /* Accessors */
@@ -200,6 +201,8 @@ class MemoryBankShiftRegisterBanks {
     void build_wl_port_fast_lookup() const;
 
   private: /* Internal data */
+    vtr::vector<ConfigRegionId, ConfigRegionId> config_region_ids_;
+
     /* General information about the BL shift register bank */
     vtr::vector<ConfigRegionId, vtr::vector<FabricBitLineBankId, FabricBitLineBankId>> bl_bank_ids_;
     vtr::vector<ConfigRegionId, vtr::vector<FabricBitLineBankId, std::vector<BasicPort>>> bl_bank_data_ports_;
diff --git a/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.cpp b/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.cpp
index 51be3783a..9854398dd 100644
--- a/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.cpp
+++ b/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.cpp
@@ -244,6 +244,7 @@ int write_memory_bank_shift_register_fabric_bitstream_to_text_file(std::fstream&
                                                                    const bool& fast_configuration,
                                                                    const bool& bit_value_to_skip,
                                                                    const FabricBitstream& fabric_bitstream,
+                                                                   const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                    const bool& keep_dont_care_bits) {
   int status = 0;
 
@@ -251,7 +252,7 @@ int write_memory_bank_shift_register_fabric_bitstream_to_text_file(std::fstream&
   if (keep_dont_care_bits) {
     dont_care_bit = DONT_CARE_CHAR; 
   }
-  MemoryBankShiftRegisterFabricBitstream fabric_bits = build_memory_bank_shift_register_fabric_bitstream(fabric_bitstream, fast_configuration, bit_value_to_skip, dont_care_bit);
+  MemoryBankShiftRegisterFabricBitstream fabric_bits = build_memory_bank_shift_register_fabric_bitstream(fabric_bitstream, blwl_sr_banks, fast_configuration, bit_value_to_skip, dont_care_bit);
 
   /* Output information about how to intepret the bitstream */
   fp << "// Bitstream word count: " << fabric_bits.num_words() << std::endl;
@@ -363,6 +364,7 @@ int write_frame_based_fabric_bitstream_to_text_file(std::fstream& fp,
  *******************************************************************/
 int write_fabric_bitstream_to_text_file(const BitstreamManager& bitstream_manager,
                                         const FabricBitstream& fabric_bitstream,
+                                        const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                         const ConfigProtocol& config_protocol,
                                         const FabricGlobalPortInfo& global_ports,
                                         const std::string& fname,
@@ -439,6 +441,7 @@ int write_fabric_bitstream_to_text_file(const BitstreamManager& bitstream_manage
                                                                               apply_fast_configuration,
                                                                               bit_value_to_skip,
                                                                               fabric_bitstream,
+                                                                              blwl_sr_banks,
                                                                               keep_dont_care_bits);
     }
     break;
diff --git a/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.h b/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.h
index 6a30cf773..e63cea5c9 100644
--- a/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.h
+++ b/openfpga/src/fpga_bitstream/write_text_fabric_bitstream.h
@@ -9,6 +9,7 @@
 #include "bitstream_manager.h"
 #include "fabric_bitstream.h"
 #include "config_protocol.h"
+#include "memory_bank_shift_register_banks.h"
 #include "fabric_global_port_info.h"
 
 /********************************************************************
@@ -20,6 +21,7 @@ namespace openfpga {
 
 int write_fabric_bitstream_to_text_file(const BitstreamManager& bitstream_manager,
                                         const FabricBitstream& fabric_bitstream,
+                                        const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                         const ConfigProtocol& config_protocol,
                                         const FabricGlobalPortInfo& global_ports,
                                         const std::string& fname,
diff --git a/openfpga/src/fpga_verilog/verilog_api.cpp b/openfpga/src/fpga_verilog/verilog_api.cpp
index 8b6603fc5..b3cc6e858 100644
--- a/openfpga/src/fpga_verilog/verilog_api.cpp
+++ b/openfpga/src/fpga_verilog/verilog_api.cpp
@@ -149,6 +149,7 @@ void fpga_fabric_verilog(ModuleManager &module_manager,
 int fpga_verilog_full_testbench(const ModuleManager &module_manager,
                                 const BitstreamManager &bitstream_manager,
                                 const FabricBitstream &fabric_bitstream,
+                                const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                 const AtomContext &atom_ctx,
                                 const PlacementContext &place_ctx,
                                 const PinConstraints& pin_constraints,
@@ -175,7 +176,7 @@ int fpga_verilog_full_testbench(const ModuleManager &module_manager,
   /* Generate full testbench for verification, including configuration phase and operating phase */
   std::string top_testbench_file_path = src_dir_path + netlist_name + std::string(AUTOCHECK_TOP_TESTBENCH_VERILOG_FILE_POSTFIX);
   print_verilog_full_testbench(module_manager,
-                              bitstream_manager, fabric_bitstream,
+                              bitstream_manager, fabric_bitstream, blwl_sr_banks,
                               circuit_lib,
                               config_protocol,
                               fabric_global_port_info,
diff --git a/openfpga/src/fpga_verilog/verilog_api.h b/openfpga/src/fpga_verilog/verilog_api.h
index 454921e26..6f83d3684 100644
--- a/openfpga/src/fpga_verilog/verilog_api.h
+++ b/openfpga/src/fpga_verilog/verilog_api.h
@@ -48,6 +48,7 @@ void fpga_fabric_verilog(ModuleManager& module_manager,
 int fpga_verilog_full_testbench(const ModuleManager& module_manager,
                                 const BitstreamManager& bitstream_manager, 
                                 const FabricBitstream& fabric_bitstream, 
+                                const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                 const AtomContext& atom_ctx, 
                                 const PlacementContext& place_ctx, 
                                 const PinConstraints& pin_constraints,
diff --git a/openfpga/src/fpga_verilog/verilog_top_testbench.cpp b/openfpga/src/fpga_verilog/verilog_top_testbench.cpp
index f45e84c9d..af50ce55d 100644
--- a/openfpga/src/fpga_verilog/verilog_top_testbench.cpp
+++ b/openfpga/src/fpga_verilog/verilog_top_testbench.cpp
@@ -1132,6 +1132,7 @@ int print_verilog_top_testbench_configuration_protocol_stimulus(std::fstream& fp
                                                                 const bool& fast_configuration,
                                                                 const bool& bit_value_to_skip,
                                                                 const FabricBitstream& fabric_bitstream,
+                                                                const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                 const float& prog_clock_period,
                                                                 const float& timescale) {
   /* Validate the file stream */
@@ -1147,7 +1148,8 @@ int print_verilog_top_testbench_configuration_protocol_stimulus(std::fstream& fp
     return print_verilog_top_testbench_configuration_protocol_ql_memory_bank_stimulus(fp,
                                                                                       config_protocol, sim_settings, 
                                                                                       module_manager, top_module,
-                                                                                      fast_configuration, bit_value_to_skip, fabric_bitstream,
+                                                                                      fast_configuration, bit_value_to_skip,
+                                                                                      fabric_bitstream, blwl_sr_banks,
                                                                                       prog_clock_period, timescale);
     break;
   case CONFIG_MEM_MEMORY_BANK:
@@ -1738,7 +1740,8 @@ void print_verilog_full_testbench_bitstream(std::fstream& fp,
                                             const ModuleManager& module_manager,
                                             const ModuleId& top_module,
                                             const BitstreamManager& bitstream_manager,
-                                            const FabricBitstream& fabric_bitstream) {
+                                            const FabricBitstream& fabric_bitstream,
+                                            const MemoryBankShiftRegisterBanks& blwl_sr_banks) {
 
   /* Branch on the type of configuration protocol */
   switch (config_protocol.type()) {
@@ -1771,7 +1774,7 @@ void print_verilog_full_testbench_bitstream(std::fstream& fp,
                                                           fast_configuration, 
                                                           bit_value_to_skip,
                                                           module_manager, top_module,
-                                                          fabric_bitstream);
+                                                          fabric_bitstream, blwl_sr_banks);
     break;
   case CONFIG_MEM_FRAME_BASED:
     print_verilog_full_testbench_frame_decoder_bitstream(fp, bitstream_file,
@@ -1893,6 +1896,7 @@ void print_verilog_top_testbench_check(std::fstream& fp,
 int print_verilog_full_testbench(const ModuleManager& module_manager,
                                  const BitstreamManager& bitstream_manager,
                                  const FabricBitstream& fabric_bitstream,
+                                 const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                  const CircuitLibrary& circuit_lib,
                                  const ConfigProtocol& config_protocol,
                                  const FabricGlobalPortInfo& global_ports,
@@ -1984,7 +1988,8 @@ int print_verilog_full_testbench(const ModuleManager& module_manager,
   status = print_verilog_top_testbench_configuration_protocol_stimulus(fp, 
                                                                        config_protocol, simulation_parameters,
                                                                        module_manager, top_module,
-                                                                       fast_configuration, bit_value_to_skip, fabric_bitstream,
+                                                                       fast_configuration, bit_value_to_skip,
+                                                                       fabric_bitstream, blwl_sr_banks,
                                                                        prog_clock_period,
                                                                        VERILOG_SIM_TIMESCALE);
 
@@ -2057,7 +2062,7 @@ int print_verilog_full_testbench(const ModuleManager& module_manager,
                                          apply_fast_configuration,
                                          bit_value_to_skip,
                                          module_manager, top_module,
-                                         bitstream_manager, fabric_bitstream);
+                                         bitstream_manager, fabric_bitstream, blwl_sr_banks);
 
   /* Add signal initialization: 
    * Bypass writing codes to files due to the autogenerated codes are very large.
diff --git a/openfpga/src/fpga_verilog/verilog_top_testbench.h b/openfpga/src/fpga_verilog/verilog_top_testbench.h
index aff108d72..bbad08125 100644
--- a/openfpga/src/fpga_verilog/verilog_top_testbench.h
+++ b/openfpga/src/fpga_verilog/verilog_top_testbench.h
@@ -17,6 +17,7 @@
 #include "fabric_global_port_info.h"
 #include "vpr_netlist_annotation.h"
 #include "simulation_setting.h"
+#include "memory_bank_shift_register_banks.h"
 #include "verilog_testbench_options.h"
 
 /********************************************************************
@@ -29,6 +30,7 @@ namespace openfpga {
 int print_verilog_full_testbench(const ModuleManager& module_manager,
                                  const BitstreamManager& bitstream_manager,
                                  const FabricBitstream& fabric_bitstream,
+                                 const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                  const CircuitLibrary& circuit_lib,
                                  const ConfigProtocol& config_protocol,
                                  const FabricGlobalPortInfo& global_ports,
diff --git a/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.cpp b/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.cpp
index c933cc939..b9eb69d96 100644
--- a/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.cpp
+++ b/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.cpp
@@ -326,6 +326,7 @@ int print_verilog_top_testbench_configuration_protocol_ql_memory_bank_stimulus(s
                                                                                const bool& fast_configuration,
                                                                                const bool& bit_value_to_skip,
                                                                                const FabricBitstream& fabric_bitstream,
+                                                                               const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                const float& prog_clock_period,
                                                                                const float& timescale) {
   ModulePortId en_port_id = module_manager.find_module_port(top_module,
@@ -350,6 +351,7 @@ int print_verilog_top_testbench_configuration_protocol_ql_memory_bank_stimulus(s
   /* Reorganize the fabric bitstream by the same address across regions */
   if (CONFIG_MEM_QL_MEMORY_BANK == config_protocol.type()) {
     MemoryBankShiftRegisterFabricBitstream fabric_bits_by_addr = build_memory_bank_shift_register_fabric_bitstream(fabric_bitstream,
+                                                                                                                   blwl_sr_banks,
                                                                                                                    fast_configuration,
                                                                                                                    bit_value_to_skip);
 
@@ -541,12 +543,14 @@ void print_verilog_full_testbench_ql_memory_bank_shift_register_bitstream(std::f
                                                                           const bool& bit_value_to_skip,
                                                                           const ModuleManager& module_manager,
                                                                           const ModuleId& top_module,
-                                                                          const FabricBitstream& fabric_bitstream) {
+                                                                          const FabricBitstream& fabric_bitstream,
+                                                                          const MemoryBankShiftRegisterBanks& blwl_sr_banks) {
   /* Validate the file stream */
   valid_file_stream(fp);
 
   /* Reorganize the fabric bitstream by the same address across regions */
   MemoryBankShiftRegisterFabricBitstream fabric_bits_by_addr = build_memory_bank_shift_register_fabric_bitstream(fabric_bitstream,
+                                                                                                                 blwl_sr_banks,
                                                                                                                  fast_configuration,
                                                                                                                  bit_value_to_skip);
 
@@ -940,7 +944,8 @@ void print_verilog_full_testbench_ql_memory_bank_bitstream(std::fstream& fp,
                                                            const bool& bit_value_to_skip,
                                                            const ModuleManager& module_manager,
                                                            const ModuleId& top_module,
-                                                           const FabricBitstream& fabric_bitstream) {
+                                                           const FabricBitstream& fabric_bitstream,
+                                                           const MemoryBankShiftRegisterBanks& blwl_sr_banks) {
   if ( (BLWL_PROTOCOL_DECODER == config_protocol.bl_protocol_type())
     && (BLWL_PROTOCOL_DECODER == config_protocol.wl_protocol_type()) ) {
     print_verilog_full_testbench_ql_memory_bank_decoder_bitstream(fp, bitstream_file,
@@ -961,7 +966,7 @@ void print_verilog_full_testbench_ql_memory_bank_bitstream(std::fstream& fp,
                                                                          fast_configuration, 
                                                                          bit_value_to_skip,
                                                                          module_manager, top_module,
-                                                                         fabric_bitstream);
+                                                                         fabric_bitstream, blwl_sr_banks);
   }
 }
 
diff --git a/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.h b/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.h
index e3a4edc7e..6bcd50903 100644
--- a/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.h
+++ b/openfpga/src/fpga_verilog/verilog_top_testbench_memory_bank.h
@@ -17,6 +17,7 @@
 #include "fabric_global_port_info.h"
 #include "vpr_netlist_annotation.h"
 #include "simulation_setting.h"
+#include "memory_bank_shift_register_banks.h"
 #include "verilog_testbench_options.h"
 
 /********************************************************************
@@ -53,6 +54,7 @@ int print_verilog_top_testbench_configuration_protocol_ql_memory_bank_stimulus(s
                                                                                const bool& fast_configuration,
                                                                                const bool& bit_value_to_skip,
                                                                                const FabricBitstream& fabric_bitstream,
+                                                                               const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                const float& prog_clock_period,
                                                                                const float& timescale);
 
@@ -67,7 +69,8 @@ void print_verilog_full_testbench_ql_memory_bank_bitstream(std::fstream& fp,
                                                            const bool& bit_value_to_skip,
                                                            const ModuleManager& module_manager,
                                                            const ModuleId& top_module,
-                                                           const FabricBitstream& fabric_bitstream);
+                                                           const FabricBitstream& fabric_bitstream,
+                                                           const MemoryBankShiftRegisterBanks& blwl_sr_banks);
 
 } /* end namespace openfpga */
 
diff --git a/openfpga/src/utils/fabric_bitstream_utils.cpp b/openfpga/src/utils/fabric_bitstream_utils.cpp
index b04a28d1c..16074cb92 100644
--- a/openfpga/src/utils/fabric_bitstream_utils.cpp
+++ b/openfpga/src/utils/fabric_bitstream_utils.cpp
@@ -12,6 +12,7 @@
 #include "vtr_log.h"
 
 /* Headers from openfpgautil library */
+#include "openfpga_reserved_words.h"
 #include "openfpga_decode.h"
 
 #include "fabric_bitstream_utils.h"
@@ -397,9 +398,109 @@ std::vector<std::string> reshape_bitstream_vectors_to_first_element(const std::v
   return rotated_vectors;
 }
 
+/** @brief Split each BL vector in a configuration region into multiple shift register banks
+ *  For example
+ *  Original vector: 1xxx010xxx1
+ *  Resulting vector (2 register register banks):
+ *   1xxx0
+ *   10xxx1 
+ */
+static 
+std::vector<std::string> redistribute_bl_vectors_to_shift_register_banks(const std::vector<std::string> bl_vectors, 
+                                                                         const MemoryBankShiftRegisterBanks& blwl_sr_banks) {
+  std::vector<std::string> multi_bank_bl_vec;
+
+  /* Resize the vector by counting the dimension */
+  /* Compute the start index of each region */
+  vtr::vector<ConfigRegionId, size_t> region_start_index;
+  region_start_index.resize(blwl_sr_banks.regions().size(), 0);
+  size_t total_num_banks = 0;
+  for (const auto& region : blwl_sr_banks.regions()) {
+    region_start_index[region] = total_num_banks; 
+    total_num_banks += blwl_sr_banks.bl_banks(region).size();
+  } 
+  multi_bank_bl_vec.resize(total_num_banks);
+
+  /* Resize each bank to be memory efficient */
+  size_t vec_start_index = 0;
+  for (const auto& region : blwl_sr_banks.regions()) {
+    for (const auto& bank : blwl_sr_banks.bl_banks(region)) {
+      size_t bank_size = blwl_sr_banks.bl_bank_size(region, bank);
+      multi_bank_bl_vec[vec_start_index].resize(bank_size);
+      vec_start_index++;
+    }
+  }
+
+  for (const std::string& region_bl_vec : bl_vectors) {
+    ConfigRegionId region = ConfigRegionId(&region_bl_vec - &bl_vectors[0]);
+    for (size_t ibit = 0; ibit < region_bl_vec.size(); ++ibit) {
+      /* Find the shift register bank id and the offset in data lines */
+      BasicPort bl_port(std::string(MEMORY_BL_PORT_NAME), ibit, ibit);
+      FabricBitLineBankId bank_id = blwl_sr_banks.find_bl_shift_register_bank_id(region, bl_port); 
+      BasicPort sr_port = blwl_sr_banks.find_bl_shift_register_bank_data_port(region, bl_port); 
+      VTR_ASSERT(1 == sr_port.get_width());
+      
+      size_t vec_index = region_start_index[region] + size_t(bank_id);
+      multi_bank_bl_vec[vec_index][sr_port.get_lsb()] = region_bl_vec[ibit];
+    }
+  }
+
+  return multi_bank_bl_vec;
+}
+
+/** @brief Split each WL vector in a configuration region into multiple shift register banks
+ *  For example
+ *  Original vector: 1xxx010xxx1
+ *  Resulting vector (2 register register banks):
+ *   1xxx0
+ *   10xxx1 
+ */
+static 
+std::vector<std::string> redistribute_wl_vectors_to_shift_register_banks(const std::vector<std::string> wl_vectors, 
+                                                                         const MemoryBankShiftRegisterBanks& blwl_sr_banks) {
+  std::vector<std::string> multi_bank_wl_vec;
+
+  /* Resize the vector by counting the dimension */
+  /* Compute the start index of each region */
+  vtr::vector<ConfigRegionId, size_t> region_start_index;
+  region_start_index.resize(blwl_sr_banks.regions().size(), 0);
+  size_t total_num_banks = 0;
+  for (const auto& region : blwl_sr_banks.regions()) {
+    region_start_index[region] = total_num_banks; 
+    total_num_banks += blwl_sr_banks.wl_banks(region).size();
+  } 
+  multi_bank_wl_vec.resize(total_num_banks);
+
+  /* Resize each bank to be memory efficient */
+  size_t vec_start_index = 0;
+  for (const auto& region : blwl_sr_banks.regions()) {
+    for (const auto& bank : blwl_sr_banks.wl_banks(region)) {
+      size_t bank_size = blwl_sr_banks.wl_bank_size(region, bank);
+      multi_bank_wl_vec[vec_start_index].resize(bank_size);
+      vec_start_index++;
+    }
+  }
+
+  for (const std::string& region_wl_vec : wl_vectors) {
+    ConfigRegionId region = ConfigRegionId(&region_wl_vec - &wl_vectors[0]);
+    for (size_t ibit = 0; ibit < region_wl_vec.size(); ++ibit) {
+      /* Find the shift register bank id and the offset in data lines */
+      BasicPort wl_port(std::string(MEMORY_WL_PORT_NAME), ibit, ibit);
+      FabricWordLineBankId bank_id = blwl_sr_banks.find_wl_shift_register_bank_id(region, wl_port); 
+      BasicPort sr_port = blwl_sr_banks.find_wl_shift_register_bank_data_port(region, wl_port); 
+      VTR_ASSERT(1 == sr_port.get_width());
+      
+      size_t vec_index = region_start_index[region] + size_t(bank_id);
+      multi_bank_wl_vec[vec_index][sr_port.get_lsb()] = region_wl_vec[ibit];
+    }
+  }
+
+  return multi_bank_wl_vec;
+}
+
 MemoryBankShiftRegisterFabricBitstream build_memory_bank_shift_register_fabric_bitstream(const FabricBitstream& fabric_bitstream,
+                                                                                         const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                          const bool& fast_configuration,
-                                                                                         //const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                          const bool& bit_value_to_skip,
                                                                                          const char& dont_care_bit) {
   MemoryBankFlattenFabricBitstream raw_fabric_bits = build_memory_bank_flatten_fabric_bitstream(fabric_bitstream, fast_configuration, bit_value_to_skip, dont_care_bit);
@@ -411,7 +512,10 @@ MemoryBankShiftRegisterFabricBitstream build_memory_bank_shift_register_fabric_b
 
     MemoryBankShiftRegisterFabricBitstreamWordId word_id = fabric_bits.create_word();
 
-    std::vector<std::string> reshaped_bl_vectors = reshape_bitstream_vectors_to_first_element(bl_vec, dont_care_bit);
+    /* Redistribute the BL vector to multiple banks */
+    std::vector<std::string> multi_bank_bl_vec = redistribute_bl_vectors_to_shift_register_banks(bl_vec, blwl_sr_banks);
+
+    std::vector<std::string> reshaped_bl_vectors = reshape_bitstream_vectors_to_first_element(multi_bank_bl_vec, dont_care_bit);
     /* Reverse the vectors due to the shift register chain nature: first-in first-out */
     std::reverse(reshaped_bl_vectors.begin(), reshaped_bl_vectors.end());
     /* Add the BL word to final bitstream */
@@ -419,7 +523,10 @@ MemoryBankShiftRegisterFabricBitstream build_memory_bank_shift_register_fabric_b
       fabric_bits.add_bl_vectors(word_id, reshaped_bl_vec); 
     }
 
-    std::vector<std::string> reshaped_wl_vectors = reshape_bitstream_vectors_to_first_element(wl_vec, dont_care_bit);
+    /* Redistribute the WL vector to multiple banks */
+    std::vector<std::string> multi_bank_wl_vec = redistribute_wl_vectors_to_shift_register_banks(wl_vec, blwl_sr_banks);
+
+    std::vector<std::string> reshaped_wl_vectors = reshape_bitstream_vectors_to_first_element(multi_bank_wl_vec, dont_care_bit);
     /* Reverse the vectors due to the shift register chain nature: first-in first-out */
     std::reverse(reshaped_wl_vectors.begin(), reshaped_wl_vectors.end());
     /* Add the BL word to final bitstream */
diff --git a/openfpga/src/utils/fabric_bitstream_utils.h b/openfpga/src/utils/fabric_bitstream_utils.h
index ac0bca14a..3cb1285bb 100644
--- a/openfpga/src/utils/fabric_bitstream_utils.h
+++ b/openfpga/src/utils/fabric_bitstream_utils.h
@@ -95,8 +95,8 @@ MemoryBankFlattenFabricBitstream build_memory_bank_flatten_fabric_bitstream(cons
  * @note the std::map may cause large memory footprint for large bitstream databases!
  *******************************************************************/
 MemoryBankShiftRegisterFabricBitstream build_memory_bank_shift_register_fabric_bitstream(const FabricBitstream& fabric_bitstream,
+                                                                                         const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                          const bool& fast_configuration,
-                                                                                         //const MemoryBankShiftRegisterBanks& blwl_sr_banks,
                                                                                          const bool& bit_value_to_skip,
                                                                                          const char& dont_care_bit = 'x');