From f50d7a420a5ac4a4788a5cb1fb891663810c5f6f Mon Sep 17 00:00:00 2001
From: Philippe Tillet <ptillet@g.harvard.edu>
Date: Mon, 15 Jul 2019 21:16:50 -0700
Subject: [PATCH] [runtime/jit] fixed bug in multi-threaded auto-tuning

---
 lib/dnn/shift.cpp   | 2 +-
 lib/runtime/jit.cpp | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/lib/dnn/shift.cpp b/lib/dnn/shift.cpp
index 87aaf32cb..d06366550 100644
--- a/lib/dnn/shift.cpp
+++ b/lib/dnn/shift.cpp
@@ -317,7 +317,7 @@ const tunable int32 TM = {16, 32, 64, 128};
 const tunable int32 TN = {16, 32, 64, 128};
 const tunable int32 TK = {)" + std::to_string(TK_) + "};";
 if(op_ == WGRAD)
-  result += "const tunable int32 GZ = {1, 4, 16};";
+  result += "const tunable int32 GZ = {1};";
 else
   result += "const tunable int32 GZ = {1};";
 
diff --git a/lib/runtime/jit.cpp b/lib/runtime/jit.cpp
index 15d33b029..fa970653a 100644
--- a/lib/runtime/jit.cpp
+++ b/lib/runtime/jit.cpp
@@ -43,8 +43,8 @@ void loop_nest(std::vector<size_t> const & ranges,
   size_t current = 0;
   while(true){
     //Execute function
-//    pool.add_job([values, &f](){ f(values); });
-    f(values);
+    pool.add_job([values, &f](){ f(values); });
+//    f(values);
     //Increment counters
     while(values[i]++ == ranges[i] - 1){
       if(i == 0)
@@ -210,9 +210,9 @@ jit::tune_res_t jit::autotune(const char *name, const char *src, benchmark_t ben
     std::unique_ptr<driver::module> module(driver::module::create(driver_context_, &*ll_module));
     std::unique_ptr<driver::kernel> kernel(driver::kernel::create(module.get(), name));
     double perf;
-    perf = benchmark(kernel.get(), info);
     {
       std::lock_guard<std::mutex> lock(mutex);
+      perf = benchmark(kernel.get(), info);
       if(perf > best.perf){
         best.perf = perf;
         best.params = params;