Add number of threads setting for MemlessPoly

8 files changed, 41 insertions, 18 deletions

diff --git a/dpd/dpd.ini b/dpd/dpd.ini
index 08564d9..21d7d45 100644
--- a/dpd/dpd.ini
+++ b/dpd/dpd.ini
@@ -24,6 +24,10 @@ enabled=1
 enabled=1
 polycoeffile=dpdpoly.coef
 
+# How many threads to use for the predistorter.
+# If not set, detect automatically.
+#num_threads=2
+
 [output]
 # to prepare a file for the dpd/iq_file_server.py script,
 # use output=file
diff --git a/src/ConfigParser.cpp b/src/ConfigParser.cpp
index 459811f..f27eb08 100644
--- a/src/ConfigParser.cpp
+++ b/src/ConfigParser.cpp
@@ -172,6 +172,9 @@ static void parse_configfile(
     if (pt.get("poly.enabled", 0) == 1) {
         mod_settings.polyCoefFilename =
             pt.get<std::string>("poly.polycoeffile", "default");
+
+        mod_settings.polyNumThreads =
+            pt.get<int>("poly.num_threads", 0);
     }
 
     // Output options
diff --git a/src/ConfigParser.h b/src/ConfigParser.h
index 22a4fc5..89f0fb7 100644
--- a/src/ConfigParser.h
+++ b/src/ConfigParser.h
@@ -75,7 +75,7 @@ struct mod_settings_t {
     std::string filterTapsFilename = "";
 
     std::string polyCoefFilename = "";
-
+    unsigned polyNumThreads = 0;
 
 #if defined(HAVE_OUTPUT_UHD)
     OutputUHDConfig outputuhd_conf;
diff --git a/src/DabMod.cpp b/src/DabMod.cpp
index 7c342a2..1f10fb8 100644
--- a/src/DabMod.cpp
+++ b/src/DabMod.cpp
@@ -328,7 +328,8 @@ int launch_modulator(int argc, char* argv[])
                 mod_settings.normalise,
                 mod_settings.gainmodeVariance,
                 mod_settings.filterTapsFilename,
-                mod_settings.polyCoefFilename);
+                mod_settings.polyCoefFilename,
+                mod_settings.polyNumThreads);
 
         if (format_converter) {
             flowgraph.connect(modulator, format_converter);
@@ -433,7 +434,8 @@ int launch_modulator(int argc, char* argv[])
                     mod_settings.normalise,
                     mod_settings.gainmodeVariance,
                     mod_settings.filterTapsFilename,
-                    mod_settings.polyCoefFilename);
+                    mod_settings.polyCoefFilename,
+                    mod_settings.polyNumThreads);
 
             if (format_converter) {
                 flowgraph.connect(modulator, format_converter);
diff --git a/src/DabModulator.cpp b/src/DabModulator.cpp
index 5282a2d..ccc2085 100644
--- a/src/DabModulator.cpp
+++ b/src/DabModulator.cpp
@@ -63,7 +63,8 @@ DabModulator::DabModulator(
         float& digGain, float normalise,
         float gainmodeVariance,
         const std::string& filterTapsFilename,
-        const std::string& polyCoefFilename
+        const std::string& polyCoefFilename,
+        unsigned int polyNumThreads
         ) :
     ModInput(),
     myOutputRate(outputRate),
@@ -77,6 +78,7 @@ DabModulator::DabModulator(
     myFlowgraph(NULL),
     myFilterTapsFilename(filterTapsFilename),
     myPolyCoefFilename(polyCoefFilename),
+    myPolyNumThreads(polyNumThreads),
     myTiiConfig(tiiConfig)
 {
     PDEBUG("DabModulator::DabModulator(%u, %u, %u, %zu) @ %p\n",
@@ -222,7 +224,8 @@ int DabModulator::process(Buffer* dataOut)
 
         shared_ptr<MemlessPoly> cifPoly;
         if (not myPolyCoefFilename.empty()) {
-            cifPoly = make_shared<MemlessPoly>(myPolyCoefFilename);
+            cifPoly = make_shared<MemlessPoly>(
+                    myPolyCoefFilename, myPolyNumThreads);
             rcs.enrol(cifPoly.get());
         }
 
diff --git a/src/DabModulator.h b/src/DabModulator.h
index 0c691dd..56a6f91 100644
--- a/src/DabModulator.h
+++ b/src/DabModulator.h
@@ -56,7 +56,9 @@ public:
             float& digGain, float normalise,
             float gainmodeVariance,
             const std::string& filterTapsFilename,
-            const std::string& polyCoefFilename);
+            const std::string& polyCoefFilename,
+            unsigned int polyNumThreads);
+
     DabModulator(const DabModulator& other) = delete;
     DabModulator& operator=(const DabModulator& other) = delete;
     virtual ~DabModulator();
@@ -82,6 +84,7 @@ protected:
     OutputMemory* myOutput;
     std::string myFilterTapsFilename;
     std::string myPolyCoefFilename;
+    unsigned int myPolyNumThreads;
     tii_config_t& myTiiConfig;
 
     size_t myNbSymbols;
diff --git a/src/MemlessPoly.cpp b/src/MemlessPoly.cpp
index 71ceac3..b0d950c 100644
--- a/src/MemlessPoly.cpp
+++ b/src/MemlessPoly.cpp
@@ -53,9 +53,10 @@ static const std::array<complexf, 8> default_coefficients({{
         }});
 
 
-MemlessPoly::MemlessPoly(const std::string& coefs_file) :
+MemlessPoly::MemlessPoly(const std::string& coefs_file, unsigned int num_threads) :
     PipelinedModCodec(),
     RemoteControllable("memlesspoly"),
+    m_num_threads(num_threads),
     m_coefs(),
     m_coefs_file(coefs_file),
     m_coefs_mutex()
@@ -63,6 +64,16 @@ MemlessPoly::MemlessPoly(const std::string& coefs_file) :
     PDEBUG("MemlessPoly::MemlessPoly(%s) @ %p\n",
             coefs_file.c_str(), this);
 
+    if (m_num_threads == 0) {
+        const unsigned int hw_concurrency = std::thread::hardware_concurrency();
+        etiLog.level(info) << "Polynomial Predistorter will use " <<
+            hw_concurrency << " threads (auto detected)";
+    }
+    else {
+        etiLog.level(info) << "Polynomial Predistorter will use " <<
+            m_num_threads << " threads (set in config file)";
+    }
+
     RC_ADD_PARAMETER(ncoefs, "(Read-only) number of coefficients.");
     RC_ADD_PARAMETER(coeffile, "Filename containing coefficients. When written to, the new file gets automatically loaded.");
 
@@ -156,12 +167,15 @@ 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();
 
-        if (hw_concurrency) {
-            const size_t step = sizeOut / hw_concurrency;
+        const unsigned int num_threads =
+            (m_num_threads > 0) ? m_num_threads : hw_concurrency;
+
+        if (num_threads) {
+            const size_t step = sizeOut / num_threads;
             vector<future<void> > flags;
 
             size_t start = 0;
-            for (size_t i = 0; i < hw_concurrency - 1; i++) {
+            for (size_t i = 0; i < num_threads - 1; i++) {
                 flags.push_back(async(launch::async, apply_coeff,
                             m_coefs, in, start, start + step, out));
 
@@ -177,13 +191,6 @@ int MemlessPoly::internal_process(Buffer* const dataIn, Buffer* dataOut)
             }
         }
         else {
-            static bool error_printed = false;
-            if (not error_printed) {
-                etiLog.level(warn) <<
-                    "Your platform doesn't seem to have hardware concurrency. "
-                    "MemlessPoly will run single-threaded";
-            }
-            // For some reason we don't have hw concurrency.
             apply_coeff(m_coefs, in, 0, sizeOut, out);
         }
     }
diff --git a/src/MemlessPoly.h b/src/MemlessPoly.h
index 9fe19d7..b7fd81e 100644
--- a/src/MemlessPoly.h
+++ b/src/MemlessPoly.h
@@ -52,7 +52,7 @@ typedef std::complex<float> complexf;
 class MemlessPoly : public PipelinedModCodec, public RemoteControllable
 {
 public:
-    MemlessPoly(const std::string& coefs_file);
+    MemlessPoly(const std::string& coefs_file, unsigned int num_threads);
 
     virtual const char* name() { return "MemlessPoly"; }
 
@@ -67,6 +67,7 @@ private:
     int internal_process(Buffer* const dataIn, Buffer* dataOut);
     void load_coefficients(const std::string &coefFile);
 
+    unsigned int m_num_threads;
     std::vector<complexf> m_coefs;
     std::string m_coefs_file;
     mutable std::mutex m_coefs_mutex;