From d0b1a6c3cbd604a6601125d68fbfafaa64976de0 Mon Sep 17 00:00:00 2001
From: Kibaek Kim <kibaek.k@gmail.com>
Date: Tue, 8 Aug 2023 22:55:01 -0500
Subject: [PATCH 1/3] fixed warnings for race condition

---
 src/CUDA/dcauchy.jl | 2 ++
 src/CUDA/dprsrch.jl | 1 +
 src/CUDA/dspcg.jl   | 1 +
 3 files changed, 4 insertions(+)

diff --git a/src/CUDA/dcauchy.jl b/src/CUDA/dcauchy.jl
index cbff20b..2a5a14e 100644
--- a/src/CUDA/dcauchy.jl
+++ b/src/CUDA/dcauchy.jl
@@ -33,6 +33,7 @@
         q = p5*ddot(n,s,1,wa,1) + gts
         interp = (q >= mu0*gts)
     end
+    CUDA.sync_threads()
 
     # Either interpolate or extrapolate to find a successful step.
 
@@ -81,6 +82,7 @@
                 search = false
             end
         end
+        CUDA.sync_threads()
 
         # Recover the last successful step.
 
diff --git a/src/CUDA/dprsrch.jl b/src/CUDA/dprsrch.jl
index 0f645a9..89187f6 100644
--- a/src/CUDA/dprsrch.jl
+++ b/src/CUDA/dprsrch.jl
@@ -46,6 +46,7 @@
             alpha = interpf*alpha
         end
     end
+    CUDA.sync_threads()
 
     # Force at least one more constraint to be added to the active
     # set if alpha < brptmin and the full step is not successful.
diff --git a/src/CUDA/dspcg.jl b/src/CUDA/dspcg.jl
index 4895d72..605d48b 100644
--- a/src/CUDA/dspcg.jl
+++ b/src/CUDA/dspcg.jl
@@ -51,6 +51,7 @@
                 end
             end
         end
+        CUDA.sync_threads()
         nfree = CUDA.shfl_sync(0xffffffff, nfree, 1)
 
         # Exit if there are no free constraints.

From 1537a8b6e97e0cbd0303e77a8f28055df42b4ee6 Mon Sep 17 00:00:00 2001
From: Kibaek Kim <kibaek.k@gmail.com>
Date: Thu, 10 Aug 2023 14:57:15 -0500
Subject: [PATCH 2/3] revert

---
 src/CUDA/dcauchy.jl | 2 --
 src/CUDA/dprsrch.jl | 1 -
 2 files changed, 3 deletions(-)

diff --git a/src/CUDA/dcauchy.jl b/src/CUDA/dcauchy.jl
index 2a5a14e..cbff20b 100644
--- a/src/CUDA/dcauchy.jl
+++ b/src/CUDA/dcauchy.jl
@@ -33,7 +33,6 @@
         q = p5*ddot(n,s,1,wa,1) + gts
         interp = (q >= mu0*gts)
     end
-    CUDA.sync_threads()
 
     # Either interpolate or extrapolate to find a successful step.
 
@@ -82,7 +81,6 @@
                 search = false
             end
         end
-        CUDA.sync_threads()
 
         # Recover the last successful step.
 
diff --git a/src/CUDA/dprsrch.jl b/src/CUDA/dprsrch.jl
index 89187f6..0f645a9 100644
--- a/src/CUDA/dprsrch.jl
+++ b/src/CUDA/dprsrch.jl
@@ -46,7 +46,6 @@
             alpha = interpf*alpha
         end
     end
-    CUDA.sync_threads()
 
     # Force at least one more constraint to be added to the active
     # set if alpha < brptmin and the full step is not successful.

From 03c23328e32ae659faa74461cceb112ffd36b352 Mon Sep 17 00:00:00 2001
From: Kibaek Kim <kibaek.k@gmail.com>
Date: Thu, 10 Aug 2023 14:57:54 -0500
Subject: [PATCH 3/3] maybe the root cause of race warnings

---
 src/CUDA/ddot.jl | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/CUDA/ddot.jl b/src/CUDA/ddot.jl
index 0a7740b..638ca24 100644
--- a/src/CUDA/ddot.jl
+++ b/src/CUDA/ddot.jl
@@ -9,5 +9,6 @@
     @inbounds for i=1:n
         v += dx[i]*dy[i]
     end
+    CUDA.sync_threads()
     return v
 end