diff --git a/taichi/codegen/amdgpu/CMakeLists.txt b/taichi/codegen/amdgpu/CMakeLists.txt
index fa30cd7f9c736..9c5a04d589c2b 100644
--- a/taichi/codegen/amdgpu/CMakeLists.txt
+++ b/taichi/codegen/amdgpu/CMakeLists.txt
@@ -14,4 +14,4 @@ target_include_directories(amdgpu_codegen
   )
 
 target_link_libraries(amdgpu_codegen PRIVATE taichi_util)
-# target_link_libraries(amdgpu_codegen PRIVATE amdgpu_runtime)
+target_link_libraries(amdgpu_codegen PRIVATE amdgpu_runtime)
diff --git a/taichi/runtime/amdgpu/jit_amdgpu.h b/taichi/runtime/amdgpu/jit_amdgpu.h
index 9aa06627f43b6..6bf15fcbd621f 100644
--- a/taichi/runtime/amdgpu/jit_amdgpu.h
+++ b/taichi/runtime/amdgpu/jit_amdgpu.h
@@ -66,15 +66,17 @@ class JITModuleAMDGPU : public JITModule {
   }
 
   void call(const std::string &name,
-            const std::vector<void *> &arg_pointers) override {
-    launch(name, 1, 1, 0, arg_pointers);
+            const std::vector<void *> &arg_pointers,
+            const std::vector<int> &arg_sizes) override {
+    launch(name, 1, 1, 0, arg_pointers, arg_sizes);
   }
 
   void launch(const std::string &name,
               std::size_t grid_dim,
               std::size_t block_dim,
               std::size_t dynamic_shared_mem_bytes,
-              std::vector<void *> &arg_pointers) override {
+              std::vector<void *> &arg_pointers,
+              std::vector<int> &arg_sizes) override {
     auto func = lookup_function(name);
     AMDGPUContext::get_instance().launch(func, name, arg_pointers, grid_dim,
                                          block_dim, dynamic_shared_mem_bytes);
diff --git a/tests/cpp/backends/amdgpu_device_test.cpp b/tests/cpp/backends/amdgpu_device_test.cpp
index 9851d7fd4322c..6c1bc6a273dab 100644
--- a/tests/cpp/backends/amdgpu_device_test.cpp
+++ b/tests/cpp/backends/amdgpu_device_test.cpp
@@ -195,6 +195,71 @@ TEST(AMDGPU, ConvertFuncParamAddressSpacePass) {
   }
 }
 
+TEST(AMDGPU, ConvertProgramAndLaunch) {
+  std::string program = 
+"target datalayout = \"e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7\"\n"
+"target triple = \"amdgcn-amd-amdhsa\"\n"
+"define amdgpu_kernel void @runtime_add(double addrspace(1)* %0, double addrspace(1)* %1, double addrspace(1)* %2) #0 {\n"
+"  %4 = alloca double*, align 8, addrspace(5)\n"
+"  %5 = addrspacecast double addrspace(1)* %2 to double*\n"
+"  %6 = addrspacecast double addrspace(1)* %1 to double*\n"
+"  %7 = addrspacecast double addrspace(1)* %0 to double*\n"
+"  %8 = addrspacecast double* addrspace(5)* %4 to double**\n"
+"  %9 = alloca double*, align 8, addrspace(5)\n"
+"  %10 = addrspacecast double* addrspace(5)* %9 to double**\n"
+"  %11 = alloca double*, align 8, addrspace(5)\n"
+"  %12 = addrspacecast double* addrspace(5)* %11 to double**\n"
+"  store double* %7, double** %8, align 8\n"
+"  store double* %6, double** %10, align 8\n"
+"  store double* %5, double** %12, align 8\n"
+"  %13 = load double*, double** %8, align 8\n"
+"  %14 = load double, double* %13, align 8\n"
+"  %15 = load double*, double** %10, align 8\n"
+"  %16 = load double, double* %15, align 8\n"
+"  %17 = fadd contract double %14, %16\n"
+"  %18 = load double*, double** %12, align 8\n"
+"  store double %17, double* %18, align 8\n"
+"  ret void\n"
+"}\n";
+  llvm::LLVMContext llvm_context;
+  llvm::SMDiagnostic diagnostic_err;
+  std::unique_ptr<llvm::Module> llvm_module = llvm::parseIR(
+      llvm::MemoryBuffer::getMemBuffer(program)->getMemBufferRef(),
+      diagnostic_err, llvm_context);
+
+  // auto amdgpu_session = new JITSessionAMDGPU(new TaichiLLVMContext(new CompileConfig, Arch::amdgpu), new CompileConfig(), llvm::DataLayout(""));
+  LLVMInitializeAMDGPUTarget();
+  LLVMInitializeAMDGPUTargetMC();
+  LLVMInitializeAMDGPUTargetInfo();
+  LLVMInitializeAMDGPUAsmPrinter();
+  LLVMInitializeAMDGPUAsmParser();
+  auto amdgpu_session = new JITSessionAMDGPU(nullptr, new CompileConfig(), llvm::DataLayout(""));
+  auto amdgpu_module = amdgpu_session->add_module(std::move(llvm_module), 0);
+  std::vector<void *> arg_pointers;
+  std::vector<int> arg_sizes;
+  double *args[3];
+  size_t arg_size = sizeof(double);
+  AMDGPUDriver::get_instance().malloc((void**)&(args[0]), sizeof(double) * 3);
+  args[1] = args[0] + 1;
+  args[2] = args[0] + 2;
+  double a = 10.0;
+  double b = 7.0;
+  double ret;
+  AMDGPUDriver::get_instance().memcpy_host_to_device(args[0], &a, sizeof(double));
+  AMDGPUDriver::get_instance().memcpy_host_to_device(args[1], &b, sizeof(double));
+  arg_pointers.push_back((void *)&args[0]);
+  arg_pointers.push_back((void *)&args[1]);
+  arg_pointers.push_back((void *)&args[2]);
+  arg_sizes.push_back(arg_size);
+  arg_sizes.push_back(arg_size);
+  arg_sizes.push_back(arg_size);
+  amdgpu_module->call("runtime_add", arg_pointers, arg_sizes);
+  AMDGPUDriver::get_instance().stream_synchronize(nullptr);
+  AMDGPUDriver::get_instance().memcpy_device_to_host(&ret, args[2], sizeof(double));
+  EXPECT_EQ(ret, 17);
+  AMDGPUDriver::get_instance().mem_free(args[0]);
+}
+
 }  // namespace lang
 }  // namespace taichi
 #endif