diff --git a/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/Makefile b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/Makefile
new file mode 100644
index 000000000..f5c87bb1a
--- /dev/null
+++ b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/Makefile
@@ -0,0 +1,34 @@
+#########################################################
+# Network Configutaion
+# If not configured, Will use default Values
+	bsg_global_X ?= $(bsg_tiles_X)
+	bsg_global_Y ?= $(bsg_tiles_Y)+1
+
+#########################################################
+#Tile group configuration
+# If not configured, Will use default Values
+	bsg_tiles_org_X ?= 0
+	bsg_tiles_org_Y ?= 1
+
+# If not configured, Will use default Values
+	bsg_tiles_X ?= 4
+	bsg_tiles_Y ?= 4
+
+
+all: main.run
+
+
+KERNEL_NAME ?=kernel_hardware_shared_mem_load_store
+
+OBJECT_FILES=main.o kernel_hardware_shared_mem_load_store.o
+
+include ../../Makefile.include
+
+
+main.riscv: $(LINK_SCRIPT) $(OBJECT_FILES) $(SPMD_COMMON_OBJECTS) $(BSG_MANYCORE_LIB) ../../common/crt.o
+	$(RISCV_LINK) $(OBJECT_FILES) $(SPMD_COMMON_OBJECTS) -L. "-l:$(BSG_MANYCORE_LIB)" -o $@ $(RISCV_LINK_OPTS)
+
+
+main.o: Makefile
+
+include ../../../mk/Makefile.tail_rules
diff --git a/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.cpp b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.cpp
new file mode 100644
index 000000000..bbb6814bd
--- /dev/null
+++ b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.cpp
@@ -0,0 +1,76 @@
+// * This kernel loads a memory block from DRAM into hardware
+//   tile group shared memory, and stores it back to another 
+//   array to DRAM to be compared.
+// * Tile group dimensions are fixed at 4x4.
+
+// TEMPLATE_TG_DIM_X/Y must be defined before bsg_manycore.h is
+// included. bsg_tiles_X and bsg_tiles_Y must also be defined for
+// legacy reasons, but they are deprecated.
+
+
+#define TEMPLATE_TG_DIM_X 4
+#define TEMPLATE_TG_DIM_Y 4
+#define TEMPLATE_BLOCK_SIZE    1024
+#define TEMPLATE_STRIPE_SIZE   1
+#define bsg_tiles_X TEMPLATE_TG_DIM_X
+#define bsg_tiles_Y TEMPLATE_TG_DIM_Y
+
+#include <bsg_manycore.h>
+#include "kernel_hardware_shared_mem_load_store.hpp"
+#include <bsg_tile_group_barrier.hpp>
+#include "bsg_shared_mem.hpp"
+
+using namespace bsg_manycore;
+
+
+bsg_barrier<bsg_tiles_X, bsg_tiles_Y> barrier;
+
+
+template <int TG_DIM_X,
+          int TG_DIM_Y,
+          int BLOCK_SIZE,
+          int STRIPE_SIZE,
+          typename T>
+    int __attribute__ ((noinline))
+    hardware_shared_mem_load_store(T *A, T *B) {
+    
+        // Declare tile-group shared memory
+        TileGroupSharedMem<T, BLOCK_SIZE, TG_DIM_X, TG_DIM_Y, STRIPE_SIZE> A_sh;
+        
+        for (int iter_x = __bsg_id; iter_x < BLOCK_SIZE; iter_x += TG_DIM_X * TG_DIM_Y) {
+            A_sh[iter_x] = A[iter_x];
+        }
+
+        barrier.sync();
+
+        for (int iter_x = __bsg_id; iter_x < BLOCK_SIZE; iter_x += TG_DIM_X * TG_DIM_Y) {
+            B[iter_x] = A_sh[iter_x];
+        }
+
+        barrier.sync();
+
+        return 0;
+    }
+
+
+extern "C" {
+    int  __attribute__ ((noinline)) kernel_hardware_shared_mem_load_store(float *A,
+                                                                          float *sum, 
+                                                                          uint32_t WIDTH, 
+                                                                          uint32_t block_size) {
+        int rc;
+        bsg_cuda_print_stat_kernel_start();
+
+        rc = hardware_shared_mem_load_store <TEMPLATE_TG_DIM_X,
+                                             TEMPLATE_TG_DIM_Y,
+                                             TEMPLATE_BLOCK_SIZE,
+                                             TEMPLATE_STRIPE_SIZE>  (A,
+                                                                     sum);
+
+        barrier.sync();
+
+        bsg_cuda_print_stat_kernel_end();
+
+        return rc;
+    }
+}
diff --git a/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.hpp b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.hpp
new file mode 100644
index 000000000..e58692d7c
--- /dev/null
+++ b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/kernel_hardware_shared_mem_load_store.hpp
@@ -0,0 +1,5 @@
+#ifndef __KERNEL_HARDWARE_SHARED_MEM_LOAD_STORE_HPP
+#define __KERNEL_HARDWARE_SHARED_MEM_LOAD_STORE_HPP
+#include <cstdint>
+
+#endif //__KERNEL_HARDWARE_SHARED_MEM_LOAD_STROE_HPP
diff --git a/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/main.c b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/main.c
new file mode 120000
index 000000000..24daac669
--- /dev/null
+++ b/software/spmd/bsg_cuda_lite_runtime/hardware_shared_mem_load_store/main.c
@@ -0,0 +1 @@
+../main/main.c
\ No newline at end of file