diff --git a/lib/runtime/function.cc b/lib/runtime/function.cc
index ead9a9ab4..715e20500 100644
--- a/lib/runtime/function.cc
+++ b/lib/runtime/function.cc
@@ -231,7 +231,7 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
   dce.run(module);
   vectorize.run(module);
   dce.run(module);
-//  ir::print(module, std::cout);
+  ir::print(module, std::cout);
   // generate llvm code
   llvm::LLVMContext ctx;
   std::unique_ptr<llvm::Module> llvm(new llvm::Module(module.get_name(), ctx));
diff --git a/tests/bench/dot.cc b/tests/bench/dot.cc
index fc2243bfc..646614afa 100644
--- a/tests/bench/dot.cc
+++ b/tests/bench/dot.cc
@@ -45,10 +45,10 @@ std::vector<double> do_bench(drv::stream* stream, bool AT, bool BT, int32_t M, i
   opt.defines.push_back({"TYPE", {ty}});
   opt.defines.push_back({"AT", {AT?"1":"0"}});
   opt.defines.push_back({"BT", {BT?"1":"0"}});
-  opt.defines.push_back({"TM", {"64", "128"}});
-  opt.defines.push_back({"TN", {"64", "128"}});
+  opt.defines.push_back({"TM", {"128"}});
+  opt.defines.push_back({"TN", {"128"}});
   opt.defines.push_back({"TK", {"8"}});
-  opt.num_warps = {2, 4, 8};
+  opt.num_warps = {4};
   // create function
   rt::function function(src::dot, opt);
   // benchmark available libraries