MemlessPoly: replace async by own threadpool

It is slightly faster, but the main advantage is that there
is no slowdown when running under gdb because it doesn't create
and destroy threads for every frame.

1 files changed, 64 insertions, 23 deletions

diff --git a/src/MemlessPoly.cpp b/src/MemlessPoly.cpp
index d5188f2..d7f9a96 100644
--- a/src/MemlessPoly.cpp
+++ b/src/MemlessPoly.cpp
@@ -53,7 +53,6 @@ using namespace std;
 MemlessPoly::MemlessPoly(const std::string& coefs_file, unsigned int num_threads) :
     PipelinedModCodec(),
     RemoteControllable("memlesspoly"),
-    m_num_threads(num_threads),
     m_coefs_am(),
     m_coefs_pm(),
     m_coefs_file(coefs_file),
@@ -62,19 +61,37 @@ MemlessPoly::MemlessPoly(const std::string& coefs_file, unsigned int num_threads
     PDEBUG("MemlessPoly::MemlessPoly(%s) @ %p\n",
             coefs_file.c_str(), this);
 
-    if (m_num_threads == 0) {
+    RC_ADD_PARAMETER(ncoefs, "(Read-only) number of coefficients.");
+    RC_ADD_PARAMETER(coeffile, "Filename containing coefficients. "
+            "When set, the file gets loaded.");
+
+    if (num_threads == 0) {
         const unsigned int hw_concurrency = std::thread::hardware_concurrency();
         etiLog.level(info) << "Polynomial Predistorter will use " <<
             hw_concurrency << " threads (auto detected)";
+
+        for (size_t i = 0; i < hw_concurrency; i++) {
+            m_workers.emplace_back();
+        }
+
+        for (auto& worker : m_workers) {
+            worker.thread = std::thread(
+                    &MemlessPoly::worker_thread, &worker);
+        }
     }
     else {
         etiLog.level(info) << "Polynomial Predistorter will use " <<
-            m_num_threads << " threads (set in config file)";
-    }
+            num_threads << " threads (set in config file)";
 
-    RC_ADD_PARAMETER(ncoefs, "(Read-only) number of coefficients.");
-    RC_ADD_PARAMETER(coeffile, "Filename containing coefficients. "
-            "When set, the file gets loaded.");
+        for (size_t i = 0; i < num_threads; i++) {
+            m_workers.emplace_back();
+        }
+
+        for (auto& worker : m_workers) {
+            worker.thread = std::thread(
+                    &MemlessPoly::worker_thread, &worker);
+        }
+    }
 
     load_coefficients(m_coefs_file);
 
@@ -100,7 +117,7 @@ void MemlessPoly::load_coefficients(const std::string &coefFile)
                 std::to_string(n_coefs) + " expected " + std::to_string(NUM_COEFS));
     }
 
-    const size_t n_entries = 2 * n_coefs;
+    const int n_entries = 2 * n_coefs;
 
     etiLog.log(debug, "MemlessPoly: Reading %d coefs...", n_entries);
 
@@ -138,7 +155,7 @@ void MemlessPoly::load_coefficients(const std::string &coefFile)
  * instead, and this allows the compiler to auto-vectorize the loop.
  */
 static void apply_coeff(
-        const vector<float> &coefs_am, const vector<float> &coefs_pm,
+        const float *__restrict coefs_am, const float *__restrict coefs_pm,
         const complexf *__restrict in, size_t start, size_t stop,
         complexf *__restrict out)
 {
@@ -178,6 +195,24 @@ static void apply_coeff(
     }
 }
 
+void MemlessPoly::worker_thread(MemlessPoly::worker_t *workerdata)
+{
+    while (true) {
+        worker_t::input_data_t in_data;
+        workerdata->in_queue.wait_and_pop(in_data);
+
+        if (in_data.terminate) {
+            break;
+        }
+
+        apply_coeff(in_data.coefs_am, in_data.coefs_pm,
+                in_data.in, in_data.start, in_data.stop,
+                in_data.out);
+
+        workerdata->out_queue.push(1);
+    }
+}
+
 int MemlessPoly::internal_process(Buffer* const dataIn, Buffer* dataOut)
 {
     dataOut->setLength(dataIn->getLength());
@@ -188,34 +223,40 @@ int MemlessPoly::internal_process(Buffer* const dataIn, Buffer* dataOut)
 
     {
         std::lock_guard<std::mutex> lock(m_coefs_mutex);
-        const unsigned int hw_concurrency = std::thread::hardware_concurrency();
-
-        const unsigned int num_threads =
-            (m_num_threads > 0) ? m_num_threads : hw_concurrency;
+        const size_t num_threads = m_workers.size();
 
-        if (num_threads) {
+        if (num_threads > 0) {
             const size_t step = sizeOut / num_threads;
-            vector<future<void> > flags;
 
             size_t start = 0;
-            for (size_t i = 0; i < num_threads - 1; i++) {
-                flags.push_back(async(launch::async, apply_coeff,
-                            m_coefs_am, m_coefs_pm,
-                            in, start, start + step, out));
+            for (auto& worker : m_workers) {
+                worker_t::input_data_t dat;
+                dat.terminate = false;
+                dat.coefs_am = m_coefs_am.data();
+                dat.coefs_pm = m_coefs_pm.data();
+                dat.in = in;
+                dat.start = start;
+                dat.stop = start + step;
+                dat.out = out;
+
+                worker.in_queue.push(dat);
 
                 start += step;
             }
 
             // Do the last in this thread
-            apply_coeff(m_coefs_am, m_coefs_pm, in, start, sizeOut, out);
+            apply_coeff(m_coefs_am.data(), m_coefs_pm.data(),
+                    in, start, sizeOut, out);
 
             // Wait for completion of the tasks
-            for (auto& f : flags) {
-                f.get();
+            for (auto& worker : m_workers) {
+                int ret;
+                worker.out_queue.wait_and_pop(ret);
             }
         }
         else {
-            apply_coeff(m_coefs_am, m_coefs_pm, in, 0, sizeOut, out);
+            apply_coeff(m_coefs_am.data(), m_coefs_pm.data(),
+                    in, 0, sizeOut, out);
         }
     }