Merge branch 'master' into next

7 files changed, 170 insertions, 22 deletions

diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc.cc b/gnuradio-core/src/lib/general/gr_fft_vcc.cc
index d07f6fa07..64dda2491 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.cc
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.cc
@@ -32,9 +32,12 @@
 #include <string.h>
 
 gr_fft_vcc_sptr
-gr_make_fft_vcc (int fft_size, bool forward,const std::vector<float> &window, bool shift)
+gr_make_fft_vcc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 bool shift, int nthreads)
 {
-  return gr_make_fft_vcc_fftw(fft_size, forward, window, shift);
+  return gr_make_fft_vcc_fftw(fft_size, forward,
+			      window, shift, nthreads);
 }
 
 gr_fft_vcc::gr_fft_vcc (const std::string &name,
@@ -62,3 +65,16 @@ gr_fft_vcc::set_window(const std::vector<float> &window)
   else 
     return false;
 }
+
+void
+gr_fft_vcc::set_nthreads(int n)
+{
+  throw std::runtime_error("gr_fft_vcc::set_nthreads not implemented.");
+}
+
+int
+gr_fft_vcc::nthreads() const
+{
+  throw std::runtime_error("gr_fft_vcc::nthreads not implemented.");
+  return 0;
+}
diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc.h b/gnuradio-core/src/lib/general/gr_fft_vcc.h
index a7c8e1162..6c3985987 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.h
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.h
@@ -30,7 +30,9 @@ class gr_fft_vcc;
 typedef boost::shared_ptr<gr_fft_vcc> gr_fft_vcc_sptr;
 
 GR_CORE_API gr_fft_vcc_sptr
-gr_make_fft_vcc (int fft_size, bool forward, const std::vector<float> &window, bool shift=false);
+gr_make_fft_vcc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 bool shift=false, int nthreads=1);
 
 /*!
  * \brief Compute forward or reverse FFT.  complex vector in / complex vector out.
@@ -42,7 +44,9 @@ class GR_CORE_API gr_fft_vcc : public gr_sync_block
 {
 protected:
   friend GR_CORE_API gr_fft_vcc_sptr
-  gr_make_fft_vcc (int fft_size, bool forward, const std::vector<float> &window, bool shift);
+  gr_make_fft_vcc (int fft_size, bool forward,
+		   const std::vector<float> &window,
+		   bool shift);
 
   unsigned int	       d_fft_size;
   std::vector<float>   d_window;
@@ -55,6 +59,9 @@ protected:
  public:
   ~gr_fft_vcc ();
 
+  virtual void set_nthreads(int n);
+  virtual int nthreads() const;
+
   bool set_window(const std::vector<float> &window);
 };
 
diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc.i b/gnuradio-core/src/lib/general/gr_fft_vcc.i
index f35316e70..0dc5353b2 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.i
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.i
@@ -23,13 +23,19 @@
 GR_SWIG_BLOCK_MAGIC(gr, fft_vcc)
 
 gr_fft_vcc_sptr 
-gr_make_fft_vcc (int fft_size, bool forward, const std::vector<float> &window, bool shift=false);
+gr_make_fft_vcc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 bool shift=false, int nthreads=1);
 
 class gr_fft_vcc : public gr_sync_block
 {
  protected:
-  gr_fft_vcc (int fft_size, bool forward, const std::vector<float> &window, bool shift);
+  gr_fft_vcc (int fft_size, bool forward,
+	      const std::vector<float> &window,
+	      bool shift);
 
  public:
   bool set_window(const std::vector<float> &window);
+  void set_nthreads(int n);
+  int nthreads() const;
 };
diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.cc b/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.cc
index 948ab20b9..e6032ad9e 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.cc
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.cc
@@ -31,16 +31,21 @@
 #include <string.h>
 
 gr_fft_vcc_sptr
-gr_make_fft_vcc_fftw (int fft_size, bool forward, const std::vector<float> &window, bool shift)
+gr_make_fft_vcc_fftw (int fft_size, bool forward,
+		      const std::vector<float> &window,
+		      bool shift, int nthreads)
 {
-  return gnuradio::get_initial_sptr(new gr_fft_vcc_fftw (fft_size, forward, window, shift));
+  return gnuradio::get_initial_sptr(new gr_fft_vcc_fftw
+				    (fft_size, forward, window,
+				     shift, nthreads));
 }
 
 gr_fft_vcc_fftw::gr_fft_vcc_fftw (int fft_size, bool forward,
-				  const std::vector<float> &window, bool shift)
+				  const std::vector<float> &window,
+				  bool shift, int nthreads)
   : gr_fft_vcc("fft_vcc_fftw", fft_size, forward, window, shift)
 {
-  d_fft = new gri_fft_complex (d_fft_size, forward);
+  d_fft = new gri_fft_complex (d_fft_size, forward, nthreads);
 }
 
 gr_fft_vcc_fftw::~gr_fft_vcc_fftw ()
@@ -48,6 +53,18 @@ gr_fft_vcc_fftw::~gr_fft_vcc_fftw ()
   delete d_fft;
 }
 
+void
+gr_fft_vcc_fftw::set_nthreads(int n)
+{
+  d_fft->set_nthreads(n);
+}
+
+int
+gr_fft_vcc_fftw::nthreads() const
+{
+  return d_fft->nthreads();
+}
+
 int
 gr_fft_vcc_fftw::work (int noutput_items,
 		  gr_vector_const_void_star &input_items,
@@ -70,7 +87,7 @@ gr_fft_vcc_fftw::work (int noutput_items,
       if(!d_forward && d_shift){
         int offset = (!d_forward && d_shift)?(d_fft_size/2):0;
         int fft_m_offset = d_fft_size - offset;
-        for (unsigned int i = 0; i < offset; i++)		// apply window
+        for (int i = 0; i < offset; i++)		// apply window
             dst[i+fft_m_offset] = in[i] * d_window[i];
         for (unsigned int i = offset; i < d_fft_size; i++)		// apply window
             dst[i-offset] = in[i] * d_window[i];
diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.h b/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.h
index 8535d133c..82b7512d7 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.h
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc_fftw.h
@@ -29,7 +29,9 @@
 class gri_fft_complex;
 
 GR_CORE_API gr_fft_vcc_sptr
-gr_make_fft_vcc_fftw (int fft_size, bool forward, const std::vector<float> &window, bool shift=false);
+gr_make_fft_vcc_fftw (int fft_size, bool forward,
+		      const std::vector<float> &window,
+		      bool shift=false, int nthreads=1);
 
 /*!
  * \brief Compute forward or reverse FFT.  complex vector in / complex vector out.
@@ -40,15 +42,22 @@ gr_make_fft_vcc_fftw (int fft_size, bool forward, const std::vector<float> &wind
 class GR_CORE_API gr_fft_vcc_fftw : public gr_fft_vcc
 {
   friend GR_CORE_API gr_fft_vcc_sptr
-  gr_make_fft_vcc_fftw (int fft_size, bool forward, const std::vector<float> &window, bool shift);
+  gr_make_fft_vcc_fftw (int fft_size, bool forward,
+			const std::vector<float> &window,
+			bool shift, int nthreads);
 
   gri_fft_complex *d_fft;
 
-  gr_fft_vcc_fftw (int fft_size, bool forward, const std::vector<float> &window, bool shift);
+  gr_fft_vcc_fftw (int fft_size, bool forward,
+		   const std::vector<float> &window,
+		   bool shift, int nthreads=1);
 
  public:
   ~gr_fft_vcc_fftw ();
 
+  void set_nthreads(int n);
+  int nthreads() const;
+
   int work (int noutput_items,
 	    gr_vector_const_void_star &input_items,
 	    gr_vector_void_star &output_items);
diff --git a/gnuradio-core/src/lib/general/gri_fft.cc b/gnuradio-core/src/lib/general/gri_fft.cc
index 0df1af25d..63e307776 100644
--- a/gnuradio-core/src/lib/general/gri_fft.cc
+++ b/gnuradio-core/src/lib/general/gri_fft.cc
@@ -78,6 +78,22 @@ gri_fftw_import_wisdom ()
 }
 
 static void
+gri_fftw_config_threading (int nthreads)
+{
+  static int fftw_threads_inited = 0;
+  
+#ifdef FFTW3F_THREADS
+  if (fftw_threads_inited == 0)
+  {
+	  fftw_threads_inited = 1;
+	  fftwf_init_threads();
+  }
+  
+  fftwf_plan_with_nthreads(nthreads);
+#endif
+}
+
+static void
 gri_fftw_export_wisdom ()
 {
   const char *filename = wisdom_filename ();
@@ -94,7 +110,7 @@ gri_fftw_export_wisdom ()
 
 // ----------------------------------------------------------------
 
-gri_fft_complex::gri_fft_complex (int fft_size, bool forward)
+gri_fft_complex::gri_fft_complex (int fft_size, bool forward, int nthreads)
 {
   // Hold global mutex during plan construction and destruction.
   gri_fft_planner::scoped_lock	lock(gri_fft_planner::mutex());
@@ -115,7 +131,10 @@ gri_fft_complex::gri_fft_complex (int fft_size, bool forward)
     throw std::runtime_error ("fftwf_malloc");
   }
 
+  d_nthreads = nthreads;
+  gri_fftw_config_threading (nthreads);
   gri_fftw_import_wisdom ();	// load prior wisdom from disk
+
   d_plan = fftwf_plan_dft_1d (fft_size,
 			      reinterpret_cast<fftwf_complex *>(d_inbuf), 
 			      reinterpret_cast<fftwf_complex *>(d_outbuf),
@@ -139,6 +158,18 @@ gri_fft_complex::~gri_fft_complex ()
   fftwf_free (d_outbuf);
 }
 
+void 
+gri_fft_complex::set_nthreads(int n)
+{
+  if (n <= 0)
+    throw std::out_of_range ("gri_fftw: invalid number of threads");
+  d_nthreads = n;
+
+#ifdef FFTW3F_THREADS
+  fftwf_plan_with_nthreads(d_nthreads);
+#endif
+}
+
 void
 gri_fft_complex::execute ()
 {
@@ -147,7 +178,7 @@ gri_fft_complex::execute ()
 
 // ----------------------------------------------------------------
 
-gri_fft_real_fwd::gri_fft_real_fwd (int fft_size)
+gri_fft_real_fwd::gri_fft_real_fwd (int fft_size, int nthreads)
 {
   // Hold global mutex during plan construction and destruction.
   gri_fft_planner::scoped_lock	lock(gri_fft_planner::mutex());
@@ -168,7 +199,10 @@ gri_fft_real_fwd::gri_fft_real_fwd (int fft_size)
     throw std::runtime_error ("fftwf_malloc");
   }
 
+  d_nthreads = nthreads;
+  gri_fftw_config_threading (nthreads);
   gri_fftw_import_wisdom ();	// load prior wisdom from disk
+  
   d_plan = fftwf_plan_dft_r2c_1d (fft_size,
 				  d_inbuf,
 				  reinterpret_cast<fftwf_complex *>(d_outbuf),
@@ -191,6 +225,18 @@ gri_fft_real_fwd::~gri_fft_real_fwd ()
   fftwf_free (d_outbuf);
 }
 
+void 
+gri_fft_real_fwd::set_nthreads(int n)
+{
+  if (n <= 0)
+    throw std::out_of_range ("gri_fftw: invalid number of threads");
+  d_nthreads = n;
+
+#ifdef FFTW3F_THREADS
+  fftwf_plan_with_nthreads(d_nthreads);
+#endif
+}
+
 void
 gri_fft_real_fwd::execute ()
 {
@@ -199,7 +245,7 @@ gri_fft_real_fwd::execute ()
 
 // ----------------------------------------------------------------
 
-gri_fft_real_rev::gri_fft_real_rev (int fft_size)
+gri_fft_real_rev::gri_fft_real_rev (int fft_size, int nthreads)
 {
   // Hold global mutex during plan construction and destruction.
   gri_fft_planner::scoped_lock	lock(gri_fft_planner::mutex());
@@ -220,11 +266,13 @@ gri_fft_real_rev::gri_fft_real_rev (int fft_size)
     throw std::runtime_error ("fftwf_malloc");
   }
 
+  d_nthreads = nthreads;
+  gri_fftw_config_threading (nthreads);
+  gri_fftw_import_wisdom ();	// load prior wisdom from disk
+  
   // FIXME If there's ever a chance that the planning functions
   // will be called in multiple threads, we've got to ensure single
   // threaded access.  They are not thread-safe.
-  
-  gri_fftw_import_wisdom ();	// load prior wisdom from disk
   d_plan = fftwf_plan_dft_c2r_1d (fft_size,
 				  reinterpret_cast<fftwf_complex *>(d_inbuf),
 				  d_outbuf,
@@ -244,6 +292,18 @@ gri_fft_real_rev::~gri_fft_real_rev ()
   fftwf_free (d_outbuf);
 }
 
+void 
+gri_fft_real_rev::set_nthreads(int n)
+{
+  if (n <= 0)
+    throw std::out_of_range ("gri_fftw: invalid number of threads");
+  d_nthreads = n;
+
+#ifdef FFTW3F_THREADS
+  fftwf_plan_with_nthreads(d_nthreads);
+#endif
+}
+
 void
 gri_fft_real_rev::execute ()
 {
diff --git a/gnuradio-core/src/lib/general/gri_fft.h b/gnuradio-core/src/lib/general/gri_fft.h
index 91a82fb55..ed80badf1 100644
--- a/gnuradio-core/src/lib/general/gri_fft.h
+++ b/gnuradio-core/src/lib/general/gri_fft.h
@@ -49,12 +49,13 @@ public:
  */
 class GR_CORE_API gri_fft_complex {
   int	      d_fft_size;
+  int         d_nthreads;
   gr_complex *d_inbuf;
   gr_complex *d_outbuf;
   void	     *d_plan;
   
 public:
-  gri_fft_complex (int fft_size, bool forward = true);
+  gri_fft_complex (int fft_size, bool forward = true, int nthreads=1);
   virtual ~gri_fft_complex ();
 
   /*
@@ -69,6 +70,16 @@ public:
   int outbuf_length () const { return d_fft_size; }
 
   /*!
+   *  Set the number of threads to use for caclulation.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   *  Get the number of threads being used by FFTW
+   */
+  int nthreads() const { return d_nthreads; }
+
+  /*!
    * compute FFT.  The input comes from inbuf, the output is placed in outbuf.
    */
   void execute ();
@@ -80,12 +91,13 @@ public:
  */
 class GR_CORE_API gri_fft_real_fwd {
   int	      d_fft_size;
+  int         d_nthreads;
   float	     *d_inbuf;
   gr_complex *d_outbuf;
   void	     *d_plan;
   
 public:
-  gri_fft_real_fwd (int fft_size);
+  gri_fft_real_fwd (int fft_size, int nthreads=1);
   virtual ~gri_fft_real_fwd ();
 
   /*
@@ -100,6 +112,16 @@ public:
   int outbuf_length () const { return d_fft_size / 2 + 1; }
 
   /*!
+   *  Set the number of threads to use for caclulation.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   *  Get the number of threads being used by FFTW
+   */
+  int nthreads() const { return d_nthreads; }
+
+  /*!
    * compute FFT.  The input comes from inbuf, the output is placed in outbuf.
    */
   void execute ();
@@ -111,12 +133,13 @@ public:
  */
 class GR_CORE_API gri_fft_real_rev {
   int	      d_fft_size;
+  int         d_nthreads;
   gr_complex *d_inbuf;
   float	     *d_outbuf;
   void	     *d_plan;
   
 public:
-  gri_fft_real_rev (int fft_size);
+  gri_fft_real_rev (int fft_size, int nthreads=1);
   virtual ~gri_fft_real_rev ();
 
   /*
@@ -131,6 +154,16 @@ public:
   int outbuf_length () const { return d_fft_size; }
 
   /*!
+   *  Set the number of threads to use for caclulation.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   *  Get the number of threads being used by FFTW
+   */
+  int nthreads() const { return d_nthreads; }
+
+  /*!
    * compute FFT.  The input comes from inbuf, the output is placed in outbuf.
    */
   void execute ();