From 002cf0d1c979857e057879d8c84b92439dbcc90d Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 22 Mar 2016 15:24:23 -0700
Subject: Use a single Barrier instead of a collection of Notifications to
 reduce the thread synchronization overhead

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

(limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h')

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index 54da77bcf..6bbf235cc 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -127,20 +127,16 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
       const Index blocksize = numext::maxi<Index>(PacketSize, (blocksz - (blocksz % PacketSize)));
       const Index numblocks = size / blocksize;
 
-      MaxSizeVector<Notification*> results(numblocks);
+      Barrier barrier(numblocks);
       for (int i = 0; i < numblocks; ++i) {
-        results.push_back(device.enqueue(&EvalRange<Evaluator, Index, Vectorizable>::run, evaluator, i*blocksize, (i+1)*blocksize));
+        device.enqueue_with_barrier(&barrier, &EvalRange<Evaluator, Index, Vectorizable>::run, evaluator, i*blocksize, (i+1)*blocksize);
       }
 
       if (numblocks * blocksize < size) {
         EvalRange<Evaluator, Index, Vectorizable>::run(evaluator, numblocks * blocksize, size);
       }
 
-      for (int i = 0; i < numblocks; ++i) {
-        wait_until_ready(results[i]);
-        delete results[i];
-      }
-
+      barrier.Wait();
     }
     evaluator.cleanup();
   }
-- 
cgit v1.2.3