From 91e94ad3a68f4743696a0dfa7e4f379c4130eacb Mon Sep 17 00:00:00 2001
From: Ruobing Han <hanruobing@gatech.edu>
Date: Thu, 15 Sep 2022 11:10:44 -0400
Subject: [PATCH] fix bug for segfault if without cudaSetDevice

---
 runtime/threadPool/src/x86/api.cpp | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/runtime/threadPool/src/x86/api.cpp b/runtime/threadPool/src/x86/api.cpp
index 0095b37..ad34b36 100644
--- a/runtime/threadPool/src/x86/api.cpp
+++ b/runtime/threadPool/src/x86/api.cpp
@@ -8,11 +8,6 @@
 #include <stdlib.h>
 #include <thread>
 
-/*
-
-
-*/
-
 /*
 Initialize the device
 */
@@ -112,6 +107,9 @@ int schedulerEnqueueKernel(cu_kernel *k) {
   Kernel Launch with numBlocks and numThreadsPerBlock
 */
 int cuLaunchKernel(cu_kernel **k) {
+  if (!device_initilized) {
+    init_device();
+  }
   // Calculate Block Size N/numBlocks
   cu_kernel *ker = *k;
   int status = C_RUN;
@@ -239,7 +237,7 @@ void cuSynchronizeBarrier() {
     init_device();
   }
   while (1) {
-    // (TODO): currently, we assume each kernel launch will  have a
+    // (TODO): currently, we assume each kernel launch has a
     // following sync
     if (scheduler->kernelQueue->size_approx() == 0) {
       int completeBlock = 0;