Print allocator and pool statistics on exit with JULIA_DEBUG set.

Author: maleadt

src/pool.jl

@@ -649,6 +649,16 @@ function __init_pool__()
   TimerOutputs.reset_timer!(alloc_to)
   TimerOutputs.reset_timer!(PoolUtils.to)
 
+  if isdebug(:init, CUDA)
+    TimerOutputs.enable_debug_timings(CUDA)
+    atexit() do
+        println("Memory pool timings:")
+        pool_timings()
+        println("Allocator timings:")
+        alloc_timings()
+    end
+  end
+
   if isinteractive()
     @async @pooled pool_cleanup()
   end

src/pool/utils.jl

@@ -103,7 +103,7 @@ export @pool_timeit
 const to = TimerOutput()
 
 macro pool_timeit(args...)
-    TimerOutputs.timer_expr(CUDA, true, :($CUDA.to), args...)
+    TimerOutputs.timer_expr(CUDA, true, :($PoolUtils.to), args...)
 end
 
 end