From d019e9cd081a0e659fd99bdc12011eeef31c456e Mon Sep 17 00:00:00 2001
From: Davis King <davis@dlib.net>
Date: Sun, 22 May 2016 15:49:40 -0400
Subject: [PATCH] Changed the trainer threading code to use dlib::thread_pool
 instead of std::async() since std::async creates new threads with each
 invocation, which in turn causes objects with thread_local storage duration
 to be reconstructed each time.  This is problematic because CUDA context
 objects for cublas and cudnn get reconstructed over and over, slowing things
 down and generally using more resources than should be used.

---
 dlib/dnn/trainer.h | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/dlib/dnn/trainer.h b/dlib/dnn/trainer.h
index dc8668032..5d6afdcc4 100644
--- a/dlib/dnn/trainer.h
+++ b/dlib/dnn/trainer.h
@@ -526,8 +526,7 @@ namespace dlib
             label_type pick_which_run_update;
             job_t next_job;
 
-            std::vector<std::future<double>> losses(devices.size());
-            std::vector<std::future<void>> update_futs(devices.size());
+            std::vector<dlib::future<double>> losses(devices.size());
 
             std::vector<tt::multi_device_tensor_averager> averagers;
             // An array of all the parameter tensors in the first network.  We will
@@ -536,6 +535,8 @@ namespace dlib
             std::vector<tensor*> reference_params;
             visit_layer_parameters(devices[0]->net, [&](size_t, tensor& t) { reference_params.push_back(&t); });
 
+            thread_pool tp(devices.size());
+
 
             size_t iteration = 0;
             while(job_pipe.dequeue(next_job))
@@ -545,7 +546,7 @@ namespace dlib
                 // right version for unsupervised or supervised training based on the type
                 // of label_type.
                 for (size_t i = 0; i < devices.size(); ++i)
-                    losses[i] = std::async(std::launch::async,[&,i](){ return compute_parameter_gradients(i, next_job, pick_which_run_update); });
+                    tp.add_task_by_value([&,i](double& loss){ loss = compute_parameter_gradients(i, next_job, pick_which_run_update); }, losses[i]);
                 // aggregate loss values from all the network computations.
                 double theloss = 0;
                 for (auto&& loss : losses)
@@ -596,10 +597,9 @@ namespace dlib
 
                 // Now apply all the updates to each device.
                 for (size_t i = 0; i < devices.size(); ++i)
-                    update_futs[i] = std::async(std::launch::async, [&,i](){ if (next_job.have_data[i]) update_parameters(i); });
+                    tp.add_task_by_value([&,i](){ if (next_job.have_data[i]) update_parameters(i); });
                 // and wait for the updates to all happen.
-                for (auto&& f : update_futs)
-                    f.wait();
+                tp.wait_for_all_tasks();
 
 
                 // Evey now and then force all the parameters to be the same just to make