Merge branch 'master' into safe_align

Conflicts:
	gnuradio-core/src/lib/filter/gr_fft_filter_ccc.cc
	gnuradio-core/src/lib/filter/gr_fft_filter_fff.cc

31 files changed, 688 insertions, 109 deletions

diff --git a/gnuradio-core/src/lib/CMakeLists.txt b/gnuradio-core/src/lib/CMakeLists.txt
index f9756feba..86f88242c 100644
--- a/gnuradio-core/src/lib/CMakeLists.txt
+++ b/gnuradio-core/src/lib/CMakeLists.txt
@@ -64,6 +64,11 @@ list(APPEND gnuradio_core_libs
     ${FFTW3F_LIBRARIES}
 )
 
+if(FFTW3F_THREADS_LIBRARIES)
+	list(APPEND gnuradio_core_libs ${FFTW3F_THREADS_LIBRARIES} )
+	add_definitions("-DFFTW3F_THREADS")
+endif()
+
 #need to link with librt on ubuntu 11.10 for shm_*
 if(LINUX)
     list(APPEND gnuradio_core_libs rt)
diff --git a/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.cc b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.cc
index d523404c9..5968e487e 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.cc
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.cc
@@ -42,13 +42,17 @@
 #include <iostream>
 #include <string.h>
 
-gr_fft_filter_ccc_sptr gr_make_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps)
+gr_fft_filter_ccc_sptr gr_make_fft_filter_ccc (int decimation,
+					       const std::vector<gr_complex> &taps,
+					       int nthreads)
 {
-  return gnuradio::get_initial_sptr(new gr_fft_filter_ccc (decimation, taps));
+  return gnuradio::get_initial_sptr(new gr_fft_filter_ccc (decimation, taps, nthreads));
 }
 
 
-gr_fft_filter_ccc::gr_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps)
+gr_fft_filter_ccc::gr_fft_filter_ccc (int decimation,
+				      const std::vector<gr_complex> &taps,
+				      int nthreads)
   : gr_sync_decimator ("fft_filter_ccc",
 		       gr_make_io_signature (1, 1, sizeof (gr_complex)),
 		       gr_make_io_signature (1, 1, sizeof (gr_complex)),
@@ -56,7 +60,13 @@ gr_fft_filter_ccc::gr_fft_filter_ccc (int decimation, const std::vector<gr_compl
     d_updated(false)
 {
   set_history(1);
-  d_filter = new gri_fft_filter_ccc_generic(decimation, taps);
+
+#if 1 // don't enable the sse version until handling it is worked out
+  d_filter = new gri_fft_filter_ccc_generic(decimation, taps, nthreads);
+#else
+  d_filter = new gri_fft_filter_ccc_sse(decimation, taps);
+#endif
+
   d_new_taps = taps;
   d_nsamples = d_filter->set_taps(taps);
   set_output_multiple(d_nsamples);
@@ -80,6 +90,23 @@ gr_fft_filter_ccc::taps () const
   return d_new_taps;
 }
 
+void
+gr_fft_filter_ccc::set_nthreads(int n)
+{
+  if(d_filter)
+    d_filter->set_nthreads(n);
+}
+
+int
+gr_fft_filter_ccc::nthreads() const
+{
+  if(d_filter)
+    return d_filter->nthreads();
+  else
+    return 0;
+}
+
+
 int
 gr_fft_filter_ccc::work (int noutput_items,
 			 gr_vector_const_void_star &input_items,
diff --git a/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.h b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.h
index 1b72a1c00..d037597e8 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.h
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.h
@@ -27,7 +27,9 @@
 
 class gr_fft_filter_ccc;
 typedef boost::shared_ptr<gr_fft_filter_ccc> gr_fft_filter_ccc_sptr;
-GR_CORE_API gr_fft_filter_ccc_sptr gr_make_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps);
+GR_CORE_API gr_fft_filter_ccc_sptr
+gr_make_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps,
+			int nthreads=1);
 
 //class gri_fft_filter_ccc_sse;
 class gri_fft_filter_ccc_generic;
@@ -39,7 +41,9 @@ class gri_fft_filter_ccc_generic;
 class GR_CORE_API gr_fft_filter_ccc : public gr_sync_decimator
 {
  private:
-  friend GR_CORE_API gr_fft_filter_ccc_sptr gr_make_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps);
+  friend GR_CORE_API gr_fft_filter_ccc_sptr
+    gr_make_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps,
+			    int nthreads);
 
   int			   d_nsamples;
   bool			   d_updated;
@@ -55,8 +59,10 @@ class GR_CORE_API gr_fft_filter_ccc : public gr_sync_decimator
    *
    * \param decimation	>= 1
    * \param taps        complex filter taps
+   * \param nthreads    number of threads for the FFT to use
    */
-  gr_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps);
+  gr_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps,
+		     int nthreads=1);
 
  public:
   ~gr_fft_filter_ccc ();
@@ -64,6 +70,16 @@ class GR_CORE_API gr_fft_filter_ccc : public gr_sync_decimator
   void set_taps (const std::vector<gr_complex> &taps);
   std::vector<gr_complex> taps () const;
 
+  /*!
+   * \brief Set number of threads to use.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   * \brief Get number of threads being used.
+   */
+  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/filter/gr_fft_filter_ccc.i b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.i
index 812920d8b..acdc347a6 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.i
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_ccc.i
@@ -24,17 +24,23 @@ GR_SWIG_BLOCK_MAGIC(gr,fft_filter_ccc)
 
 gr_fft_filter_ccc_sptr 
 gr_make_fft_filter_ccc (int decimation,
-			const std::vector<gr_complex> &taps
+			const std::vector<gr_complex> &taps,
+			int nthreads=1
 			) throw (std::invalid_argument);
 
 class gr_fft_filter_ccc : public gr_sync_decimator
 {
  private:
-  gr_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps);
+  gr_fft_filter_ccc (int decimation, const std::vector<gr_complex> &taps,
+		     int nthreads=1);
 
  public:
   ~gr_fft_filter_ccc ();
 
   void set_taps (const std::vector<gr_complex> &taps);
   std::vector<gr_complex> taps () const;
+
+  void set_nthreads(int n);
+  int nthreads() const;
+
 };
diff --git a/gnuradio-core/src/lib/filter/gr_fft_filter_fff.cc b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.cc
index 640851a1d..e4a669150 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_fff.cc
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.cc
@@ -34,13 +34,17 @@
 #include <iostream>
 #include <string.h>
 
-gr_fft_filter_fff_sptr gr_make_fft_filter_fff (int decimation, const std::vector<float> &taps)
+gr_fft_filter_fff_sptr gr_make_fft_filter_fff (int decimation,
+					       const std::vector<float> &taps,
+					       int nthreads)
 {
-  return gnuradio::get_initial_sptr(new gr_fft_filter_fff (decimation, taps));
+  return gnuradio::get_initial_sptr(new gr_fft_filter_fff (decimation, taps, nthreads));
 }
 
 
-gr_fft_filter_fff::gr_fft_filter_fff (int decimation, const std::vector<float> &taps)
+gr_fft_filter_fff::gr_fft_filter_fff (int decimation,
+				      const std::vector<float> &taps,
+				      int nthreads)
   : gr_sync_decimator ("fft_filter_fff",
 		       gr_make_io_signature (1, 1, sizeof (float)),
 		       gr_make_io_signature (1, 1, sizeof (float)),
@@ -49,7 +53,12 @@ gr_fft_filter_fff::gr_fft_filter_fff (int decimation, const std::vector<float> &
 {
   set_history(1);
   
-  d_filter = new gri_fft_filter_fff_generic(decimation, taps);
+#if 1 // don't enable the sse version until handling it is worked out
+  d_filter = new gri_fft_filter_fff_generic(decimation, taps, nthreads);
+#else
+    d_filter = new gri_fft_filter_fff_sse(decimation, taps);
+#endif
+
   d_new_taps = taps;
   d_nsamples = d_filter->set_taps(taps);
   set_output_multiple(d_nsamples);
@@ -73,6 +82,22 @@ gr_fft_filter_fff::taps () const
   return d_new_taps;
 }
 
+void
+gr_fft_filter_fff::set_nthreads(int n)
+{
+  if(d_filter)
+    d_filter->set_nthreads(n);
+}
+
+int
+gr_fft_filter_fff::nthreads() const
+{
+  if(d_filter)
+    return d_filter->nthreads();
+  else
+    return 0;
+}
+
 int
 gr_fft_filter_fff::work (int noutput_items,
 			 gr_vector_const_void_star &input_items,
diff --git a/gnuradio-core/src/lib/filter/gr_fft_filter_fff.h b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.h
index ddd8dcac2..2eeb8c646 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_fff.h
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.h
@@ -27,7 +27,9 @@
 
 class gr_fft_filter_fff;
 typedef boost::shared_ptr<gr_fft_filter_fff> gr_fft_filter_fff_sptr;
-GR_CORE_API gr_fft_filter_fff_sptr gr_make_fft_filter_fff (int decimation, const std::vector<float> &taps);
+GR_CORE_API gr_fft_filter_fff_sptr
+gr_make_fft_filter_fff (int decimation, const std::vector<float> &taps,
+			int nthreads=1);
 
 class gri_fft_filter_fff_generic;
 //class gri_fft_filter_fff_sse;
@@ -39,7 +41,9 @@ class gri_fft_filter_fff_generic;
 class GR_CORE_API gr_fft_filter_fff : public gr_sync_decimator
 {
  private:
-  friend GR_CORE_API gr_fft_filter_fff_sptr gr_make_fft_filter_fff (int decimation, const std::vector<float> &taps);
+  friend GR_CORE_API gr_fft_filter_fff_sptr
+    gr_make_fft_filter_fff (int decimation, const std::vector<float> &taps,
+			    int nthreads);
 
   int			   d_nsamples;
   bool			   d_updated;
@@ -55,15 +59,27 @@ class GR_CORE_API gr_fft_filter_fff : public gr_sync_decimator
    *
    * \param decimation	>= 1
    * \param taps        float filter taps
+   * \param nthreads    number of threads for the FFT to use
    */
-  gr_fft_filter_fff (int decimation, const std::vector<float> &taps);
-  
+  gr_fft_filter_fff (int decimation, const std::vector<float> &taps,
+		     int nthreads=1);
+
  public:
   ~gr_fft_filter_fff ();
 
   void set_taps (const std::vector<float> &taps);
   std::vector<float> taps () const;
 
+  /*!
+   * \brief Set number of threads to use.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   * \brief Get number of threads being used.
+   */
+  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/filter/gr_fft_filter_fff.i b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.i
index 7e2cde977..c8118e09e 100644
--- a/gnuradio-core/src/lib/filter/gr_fft_filter_fff.i
+++ b/gnuradio-core/src/lib/filter/gr_fft_filter_fff.i
@@ -24,17 +24,22 @@ GR_SWIG_BLOCK_MAGIC(gr,fft_filter_fff)
 
 gr_fft_filter_fff_sptr 
 gr_make_fft_filter_fff (int decimation,
-			const std::vector<float> &taps
+			const std::vector<float> &taps,
+			int nthreads=1
 			) throw (std::invalid_argument);
 
 class gr_fft_filter_fff : public gr_sync_decimator
 {
  private:
-  gr_fft_filter_fff (int decimation, const std::vector<float> &taps);
+  gr_fft_filter_fff (int decimation, const std::vector<float> &taps,
+		     int nthreads=1);
 
  public:
   ~gr_fft_filter_fff ();
 
   void set_taps (const std::vector<float> &taps);
   std::vector<float> taps () const;
+  void set_nthreads(int n);
+  int nthreads() const;
+
 };
diff --git a/gnuradio-core/src/lib/filter/gr_iir_filter_ffd.h b/gnuradio-core/src/lib/filter/gr_iir_filter_ffd.h
index 386d056e5..84be57768 100644
--- a/gnuradio-core/src/lib/filter/gr_iir_filter_ffd.h
+++ b/gnuradio-core/src/lib/filter/gr_iir_filter_ffd.h
@@ -43,16 +43,26 @@ gr_make_iir_filter_ffd (const std::vector<double> &fftaps,
  *
  * 
  * The input and output satisfy a difference equation of the form
+ \htmlonly
+ \f{
+ y[n] - \sum_{k=1}^{M} a_k y[n-k] = \sum_{k=0}^{N} b_k x[n-k]
+ \f}
+ \endhtmlonly
 
- \f[
+ \xmlonly
  y[n] - \sum_{k=1}^{M} a_k y[n-k] = \sum_{k=0}^{N} b_k x[n-k]
- \f]
+ \endxmlonly
 
  * with the corresponding rational system function
+ \htmlonly
+ \f{
+ H(z) = \ frac{\sum_{k=0}^{M} b_k z^{-k}}{1 - \sum_{k=1}^{N} a_k z^{-k}}
+ \f}
+ \endhtmlonly
 
- \f[
- H(z) = \frac{\sum_{k=0}^{M} b_k z^{-k}}{1 - \sum_{k=1}^{N} a_k z^{-k}}
- \f]
+ \xmlonly
+ H(z) = \ frac{\sum_{k=0}^{M} b_k z^{-k}}{1 - \sum_{k=1}^{N} a_k z^{-k}}
+ \endxmlonly
 
  * Note that some texts define the system function with a + in the denominator.
  * If you're using that convention, you'll need to negate the feedback taps.
diff --git a/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_cc.h b/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_cc.h
index 3ce468db2..78fa25393 100644
--- a/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_cc.h
+++ b/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_cc.h
@@ -40,16 +40,26 @@ gr_make_single_pole_iir_filter_cc (double alpha, unsigned int vlen=1);
  * \ingroup filter_blk
  *
  * The input and output satisfy a difference equation of the form
+ \htmlonly
+ \f{
+ y[n] - (1-alpha) y[n-1] = alpha x[n]
+ \f}
+ \endhtmlonly
 
- \f[
+ \xmlonly
  y[n] - (1-alpha) y[n-1] = alpha x[n]
- \f]
+ \endxmlonly
 
  * with the corresponding rational system function
-
- \f[
+ \htmlonly
+ \f{
  H(z) = \frac{alpha}{1 - (1-alpha) z^{-1}}
- \f]
+ \f}
+ \endhtmlonly
+
+ \xmlonly
+ H(z) = \ frac{alpha}{1 - (1-alpha) z^{-1}}
+ \endxmlonly
 
  * Note that some texts define the system function with a + in the denominator.
  * If you're using that convention, you'll need to negate the feedback tap.
diff --git a/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_ff.h b/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_ff.h
index d376587df..d2bae5c26 100644
--- a/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_ff.h
+++ b/gnuradio-core/src/lib/filter/gr_single_pole_iir_filter_ff.h
@@ -39,16 +39,26 @@ gr_make_single_pole_iir_filter_ff (double alpha, unsigned int vlen=1);
  * \ingroup filter_blk
  *
  * The input and output satisfy a difference equation of the form
+ \htmlonly
+ \f{
+ y[n] - (1-alpha) y[n-1] = alpha x[n]
+ \f}
+ \endhtmlonly
 
- \f[
+ \xmlonly
  y[n] - (1-alpha) y[n-1] = alpha x[n]
- \f]
+ \endxmlonly
 
  * with the corresponding rational system function
-
- \f[
+ \htmlonly
+ \f{
  H(z) = \frac{alpha}{1 - (1-alpha) z^{-1}}
- \f]
+ \f}
+ \endhtmlonly
+
+ \xmlonly
+H(z) = \ frac{alpha}{1 - (1-alpha) z^{-1}}
+ \endxmlonly
 
  * Note that some texts define the system function with a + in the denominator.
  * If you're using that convention, you'll need to negate the feedback tap.
diff --git a/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.cc b/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.cc
index d9700ad2e..c894d62aa 100644
--- a/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.cc
+++ b/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.cc
@@ -34,8 +34,9 @@
 #include <fftw3.h>
 
 gri_fft_filter_ccc_generic::gri_fft_filter_ccc_generic (int decimation, 
-							const std::vector<gr_complex> &taps)
-  : d_fftsize(-1), d_decimation(decimation), d_fwdfft(0), d_invfft(0)
+							const std::vector<gr_complex> &taps,
+							int nthreads)
+  : d_fftsize(-1), d_decimation(decimation), d_fwdfft(0), d_invfft(0), d_nthreads(nthreads)
 {
   set_taps(taps);
 }
@@ -112,12 +113,28 @@ gri_fft_filter_ccc_generic::compute_sizes(int ntaps)
   if (d_fftsize != old_fftsize){	// compute new plans
     delete d_fwdfft;
     delete d_invfft;
-    d_fwdfft = new gri_fft_complex(d_fftsize, true);
-    d_invfft = new gri_fft_complex(d_fftsize, false);
+    d_fwdfft = new gri_fft_complex(d_fftsize, true, d_nthreads);
+    d_invfft = new gri_fft_complex(d_fftsize, false, d_nthreads);
     d_xformed_taps.resize(d_fftsize);
   }
 }
 
+void
+gri_fft_filter_ccc_generic::set_nthreads(int n)
+{
+  d_nthreads = n;
+  if(d_fwdfft)
+    d_fwdfft->set_nthreads(n);
+  if(d_invfft)
+    d_invfft->set_nthreads(n);
+}
+
+int
+gri_fft_filter_ccc_generic::nthreads() const
+{
+  return d_nthreads;
+}
+
 int
 gri_fft_filter_ccc_generic::filter (int nitems, const gr_complex *input, gr_complex *output)
 {
diff --git a/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.h b/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.h
index 4db7ba50f..217b9ab83 100644
--- a/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.h
+++ b/gnuradio-core/src/lib/filter/gri_fft_filter_ccc_generic.h
@@ -42,6 +42,7 @@ class GR_CORE_API gri_fft_filter_ccc_generic
   int                      d_decimation;
   gri_fft_complex	  *d_fwdfft;		// forward "plan"
   gri_fft_complex	  *d_invfft;		// inverse "plan"
+  int                      d_nthreads;          // number of FFTW threads to use
   std::vector<gr_complex>  d_tail;		// state carried between blocks for overlap-add
   std::vector<gr_complex>  d_xformed_taps;	// Fourier xformed taps
   std::vector<gr_complex>  d_new_taps;
@@ -57,8 +58,10 @@ class GR_CORE_API gri_fft_filter_ccc_generic
    * in other blocks for complex vectors (such as gr_fft_filter_ccc).
    * \param decimation The decimation rate of the filter (int)
    * \param taps       The filter taps (complex)
+   * \param nthreads   The number of threads for the FFT to use (int)
    */
-  gri_fft_filter_ccc_generic (int decimation, const std::vector<gr_complex> &taps);
+  gri_fft_filter_ccc_generic (int decimation, const std::vector<gr_complex> &taps,
+			      int nthreads=1);
   ~gri_fft_filter_ccc_generic ();
 
   /*!
@@ -68,6 +71,16 @@ class GR_CORE_API gri_fft_filter_ccc_generic
    * \param taps       The filter taps (complex)
    */
   int set_taps (const std::vector<gr_complex> &taps);
+
+  /*!
+   * \brief Set number of threads to use.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   * \brief Get number of threads being used.
+   */
+  int nthreads() const;
   
   /*!
    * \brief Perform the filter operation
diff --git a/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.cc b/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.cc
index 64705ee5e..e7f66b714 100644
--- a/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.cc
+++ b/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.cc
@@ -33,8 +33,9 @@
 #include <cstring>
 
 gri_fft_filter_fff_generic::gri_fft_filter_fff_generic (int decimation, 
-							const std::vector<float> &taps)
-  : d_fftsize(-1), d_decimation(decimation), d_fwdfft(0), d_invfft(0)
+							const std::vector<float> &taps,
+							int nthreads)
+  : d_fftsize(-1), d_decimation(decimation), d_fwdfft(0), d_invfft(0), d_nthreads(nthreads)
 {
   set_taps(taps);
 }
@@ -105,6 +106,22 @@ gri_fft_filter_fff_generic::compute_sizes(int ntaps)
   }
 }
 
+void
+gri_fft_filter_fff_generic::set_nthreads(int n)
+{
+  d_nthreads = n;
+  if(d_fwdfft)
+    d_fwdfft->set_nthreads(n);
+  if(d_invfft)
+    d_invfft->set_nthreads(n);
+}
+
+int
+gri_fft_filter_fff_generic::nthreads() const
+{
+  return d_nthreads;
+}
+
 int
 gri_fft_filter_fff_generic::filter (int nitems, const float *input, float *output)
 {
diff --git a/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.h b/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.h
index 86658043a..be31068aa 100644
--- a/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.h
+++ b/gnuradio-core/src/lib/filter/gri_fft_filter_fff_generic.h
@@ -39,6 +39,7 @@ class GR_CORE_API gri_fft_filter_fff_generic
   int                      d_decimation;
   gri_fft_real_fwd	  *d_fwdfft;		// forward "plan"
   gri_fft_real_rev	  *d_invfft;		// inverse "plan"
+  int                      d_nthreads;          // number of FFTW threads to use
   std::vector<float>       d_tail;		// state carried between blocks for overlap-add
   std::vector<gr_complex>  d_xformed_taps;	// Fourier xformed taps
   std::vector<float>	   d_new_taps;
@@ -55,8 +56,10 @@ class GR_CORE_API gri_fft_filter_fff_generic
    * in other blocks for floating point vectors (such as gr_fft_filter_fff).
    * \param decimation The decimation rate of the filter (int)
    * \param taps       The filter taps (float)
+   * \param nthreads   The number of threads for the FFT to use (int)
    */
-  gri_fft_filter_fff_generic (int decimation, const std::vector<float> &taps);
+  gri_fft_filter_fff_generic (int decimation, const std::vector<float> &taps,
+			      int nthreads=1);
   ~gri_fft_filter_fff_generic ();
 
   /*!
@@ -68,6 +71,16 @@ class GR_CORE_API gri_fft_filter_fff_generic
   int set_taps (const std::vector<float> &taps);
   
   /*!
+   * \brief Set number of threads to use.
+   */
+  void set_nthreads(int n);
+
+  /*!
+   * \brief Get number of threads being used.
+   */
+  int nthreads() const;
+  
+  /*!
    * \brief Perform the filter operation
    *
    * \param nitems  The number of items to produce
diff --git a/gnuradio-core/src/lib/filter/qa_gri_mmse_fir_interpolator_cc.cc b/gnuradio-core/src/lib/filter/qa_gri_mmse_fir_interpolator_cc.cc
index 2fc97a78a..d0ed04238 100644
--- a/gnuradio-core/src/lib/filter/qa_gri_mmse_fir_interpolator_cc.cc
+++ b/gnuradio-core/src/lib/filter/qa_gri_mmse_fir_interpolator_cc.cc
@@ -31,6 +31,7 @@
 #include <stdio.h>
 #include <cmath>
 #include <stdexcept>
+#include <unistd.h>
 
 #define	NELEM(x) (sizeof (x) / sizeof (x[0]))
 
diff --git a/gnuradio-core/src/lib/general/gr_burst_tagger.cc b/gnuradio-core/src/lib/general/gr_burst_tagger.cc
index 4b3847b08..bd713d663 100644
--- a/gnuradio-core/src/lib/general/gr_burst_tagger.cc
+++ b/gnuradio-core/src/lib/general/gr_burst_tagger.cc
@@ -43,10 +43,39 @@ gr_burst_tagger::gr_burst_tagger(size_t itemsize)
   std::stringstream str;
   str << name() << unique_id();
 
-  d_key = pmt::pmt_string_to_symbol("burst");
+  d_true_key = pmt::pmt_string_to_symbol("burst");
+  d_true_value = pmt::PMT_T;
+  
+  d_false_key = pmt::pmt_string_to_symbol("burst");
+  d_false_value = pmt::PMT_F;
+  
   d_id  = pmt::pmt_string_to_symbol(str.str());
 }
 
+void
+gr_burst_tagger::set_true_tag (const std::string &key, bool value)
+{
+  d_true_key = pmt::pmt_string_to_symbol(key);
+  if(value == true) {
+    d_true_value = pmt::PMT_T;
+  }
+  else {
+    d_true_value = pmt::PMT_F;
+  }
+}
+
+void
+gr_burst_tagger::set_false_tag (const std::string &key, bool value)
+{
+  d_false_key = pmt::pmt_string_to_symbol(key);
+  if(value == true) {
+    d_false_value = pmt::PMT_T;
+  }
+  else {
+    d_false_value = pmt::PMT_F;
+  }
+}
+
 gr_burst_tagger::~gr_burst_tagger()
 {
 }
@@ -66,18 +95,15 @@ gr_burst_tagger::work(int noutput_items,
     if(trigger[i] > 0) {
       if(d_state == false) {
 	d_state = true;
-	pmt::pmt_t value = pmt::PMT_T;
-	add_item_tag(0, nitems_written(0)+i, d_key, value, d_id);
+	add_item_tag(0, nitems_written(0)+i, d_true_key, d_true_value, d_id);
       }
     }
     else {
       if(d_state == true) {
 	d_state = false;
-	pmt::pmt_t value = pmt::PMT_F;
-	add_item_tag(0, nitems_written(0)+i, d_key, value, d_id);
+	add_item_tag(0, nitems_written(0)+i, d_false_key, d_false_value, d_id);
       }
     }
   }
-
   return noutput_items;
 }
diff --git a/gnuradio-core/src/lib/general/gr_burst_tagger.h b/gnuradio-core/src/lib/general/gr_burst_tagger.h
index 7547ba9cc..663a146f2 100644
--- a/gnuradio-core/src/lib/general/gr_burst_tagger.h
+++ b/gnuradio-core/src/lib/general/gr_burst_tagger.h
@@ -40,14 +40,21 @@ class GR_CORE_API gr_burst_tagger : public gr_sync_block
 {
   size_t		d_itemsize;
   bool                  d_state;
-  pmt::pmt_t            d_key;
+  pmt::pmt_t            d_true_key;
+  pmt::pmt_t            d_true_value;
+  
+  pmt::pmt_t            d_false_key;
+  pmt::pmt_t            d_false_value;
+  
   pmt::pmt_t            d_id;
-
+  
   friend GR_CORE_API gr_burst_tagger_sptr gr_make_burst_tagger(size_t itemsize);
   gr_burst_tagger(size_t itemsize);
 
  public:
   ~gr_burst_tagger();
+  void set_true_tag (const std::string &key, bool value);
+  void set_false_tag (const std::string &key, bool value);
 
   int work(int noutput_items,
 	   gr_vector_const_void_star &input_items,
diff --git a/gnuradio-core/src/lib/general/gr_burst_tagger.i b/gnuradio-core/src/lib/general/gr_burst_tagger.i
index ebf1eea8c..868941fc6 100644
--- a/gnuradio-core/src/lib/general/gr_burst_tagger.i
+++ b/gnuradio-core/src/lib/general/gr_burst_tagger.i
@@ -28,4 +28,8 @@ class gr_burst_tagger : public gr_sync_block
 {
  private:
   gr_burst_tagger(size_t itemsize);
+  
+ public:
+  void set_true_tag(const std::string &key, bool value);
+  void set_false_tag(const std::string &key, bool value);
 };
diff --git a/gnuradio-core/src/lib/general/gr_fft_vcc.cc b/gnuradio-core/src/lib/general/gr_fft_vcc.cc
index d07f6fa07..f744acb93 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.cc
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.cc
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2007,2008 Free Software Foundation, Inc.
+ * Copyright 2004,2007,2008,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -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..ceabeb681 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.h
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.h
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2007,2008 Free Software Foundation, Inc.
+ * Copyright 2004,2007,2008,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -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..26d8b89a3 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vcc.i
+++ b/gnuradio-core/src/lib/general/gr_fft_vcc.i
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2007,2008,2010 Free Software Foundation, Inc.
+ * Copyright 2004,2007,2008,2010,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -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 3293e3ab8..a99beb965 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,
@@ -67,8 +84,17 @@ gr_fft_vcc_fftw::work (int noutput_items,
     
     if (d_window.size()){
       gr_complex *dst = d_fft->get_inbuf();
-      for (unsigned int i = 0; i < d_fft_size; i++)		// apply window
-	dst[i] = in[i] * d_window[i];
+      if(!d_forward && d_shift){
+        unsigned int offset = (!d_forward && d_shift)?(d_fft_size/2):0;
+        int fft_m_offset = d_fft_size - offset;
+        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];
+      } else {
+        for (unsigned int i = 0; i < d_fft_size; i++)		// apply window
+          dst[i] = in[i] * d_window[i];
+      }
     }
     else {
       if(!d_forward && d_shift) {  // apply an ifft shift on the data
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/gr_fft_vfc.cc b/gnuradio-core/src/lib/general/gr_fft_vfc.cc
index 561c63740..5fbe732dc 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vfc.cc
+++ b/gnuradio-core/src/lib/general/gr_fft_vfc.cc
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2010 Free Software Foundation, Inc.
+ * Copyright 2004,2010,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -40,12 +40,17 @@
 
 
 gr_fft_vfc_sptr
-gr_make_fft_vfc (int fft_size, bool forward, const std::vector<float> &window)
+gr_make_fft_vfc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 int nthreads)
 {
-  return gnuradio::get_initial_sptr(new gr_fft_vfc (fft_size, forward, window));
+  return gnuradio::get_initial_sptr(new gr_fft_vfc (fft_size, forward,
+						    window, nthreads));
 }
 
-gr_fft_vfc::gr_fft_vfc (int fft_size, bool forward, const std::vector<float> &window)
+gr_fft_vfc::gr_fft_vfc (int fft_size, bool forward,
+			const std::vector<float> &window,
+			int nthreads)
   : gr_sync_block ("fft_vfc",
 		   gr_make_io_signature (1, 1, fft_size * sizeof (float)),
 		   gr_make_io_signature (1, 1, fft_size * sizeof (gr_complex))),
@@ -56,7 +61,7 @@ gr_fft_vfc::gr_fft_vfc (int fft_size, bool forward, const std::vector<float> &wi
     throw std::invalid_argument ("fft_vfc: forward must == true");
   }
 
-  d_fft = new gri_fft_complex (d_fft_size, forward);
+  d_fft = new gri_fft_complex (d_fft_size, forward, nthreads);
 
   set_window(window);
 }
@@ -66,6 +71,18 @@ gr_fft_vfc::~gr_fft_vfc ()
   delete d_fft;
 }
 
+void
+gr_fft_vfc::set_nthreads(int n)
+{
+  d_fft->set_nthreads(n);
+}
+
+int
+gr_fft_vfc::nthreads() const
+{
+  return d_fft->nthreads();
+}
+
 int
 gr_fft_vfc::work (int noutput_items,
 		  gr_vector_const_void_star &input_items,
diff --git a/gnuradio-core/src/lib/general/gr_fft_vfc.h b/gnuradio-core/src/lib/general/gr_fft_vfc.h
index 6cf6b9037..84ae08f08 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vfc.h
+++ b/gnuradio-core/src/lib/general/gr_fft_vfc.h
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2010 Free Software Foundation, Inc.
+ * Copyright 2004,2010,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -32,7 +32,9 @@ class gr_fft_vfc;
 typedef boost::shared_ptr<gr_fft_vfc> gr_fft_vfc_sptr;
 
 GR_CORE_API gr_fft_vfc_sptr
-gr_make_fft_vfc (int fft_size, bool forward, const std::vector<float> &window);
+gr_make_fft_vfc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 int nthreads=1);
 
 /*!
  * \brief Compute forward FFT.  float vector in / complex vector out.
@@ -42,17 +44,24 @@ gr_make_fft_vfc (int fft_size, bool forward, const std::vector<float> &window);
 class GR_CORE_API gr_fft_vfc : public gr_sync_block
 {
   friend GR_CORE_API gr_fft_vfc_sptr
-  gr_make_fft_vfc (int fft_size, bool forward, const std::vector<float>  &window);
+  gr_make_fft_vfc (int fft_size, bool forward,
+		   const std::vector<float>  &window,
+		   int nthreads);
 
   unsigned int  d_fft_size;
   std::vector<float> d_window;
   gri_fft_complex *d_fft;
 
-  gr_fft_vfc (int fft_size, bool forward, const std::vector<float>  &window);
+  gr_fft_vfc (int fft_size, bool forward,
+	      const std::vector<float>  &window,
+	      int nthreads=1);
 
  public:
   ~gr_fft_vfc ();
 
+  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/gr_fft_vfc.i b/gnuradio-core/src/lib/general/gr_fft_vfc.i
index 149745b58..4783ae1fe 100644
--- a/gnuradio-core/src/lib/general/gr_fft_vfc.i
+++ b/gnuradio-core/src/lib/general/gr_fft_vfc.i
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2004,2010 Free Software Foundation, Inc.
+ * Copyright 2004,2010,2012 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -23,14 +23,20 @@
 GR_SWIG_BLOCK_MAGIC(gr, fft_vfc)
 
 gr_fft_vfc_sptr 
-gr_make_fft_vfc (int fft_size, bool forward, const std::vector<float> &window)
+gr_make_fft_vfc (int fft_size, bool forward,
+		 const std::vector<float> &window,
+		 int nthreads=1)
 throw(std::exception);
 
 class gr_fft_vfc : public gr_sync_block
 {
  protected:
-  gr_fft_vfc (int fft_size, bool forward, const std::vector<float> &window);
+  gr_fft_vfc (int fft_size, bool forward,
+	      const std::vector<float> &window,
+	      int nthreads=1);
   
  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/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 ();
diff --git a/gnuradio-core/src/lib/io/gr_oscope_guts.cc b/gnuradio-core/src/lib/io/gr_oscope_guts.cc
index ce7feca13..8b0d1e632 100644
--- a/gnuradio-core/src/lib/io/gr_oscope_guts.cc
+++ b/gnuradio-core/src/lib/io/gr_oscope_guts.cc
@@ -31,24 +31,36 @@
 #include <math.h>
 #include <assert.h>
 
-static const int OUTPUT_RECORD_SIZE = 2048;        // must be power of 2
+/*
+ * Bad performance if it's large, and flaky triggering if it's too small
+ */
+static const int OUTPUT_RECORD_SIZE = 1024;  // Must be power of 2
+
+/*
+ * For (slow-updated) STRIPCHART triggering, we make the record size larger, since we
+ *   potentially want to be able to "see" hours of data.  This works as long as the
+ *   update rates to a STRIPCHART are low, which they generally are--that's rather what
+ *   a stripchart is all about!
+ */
+static const int SCHART_MULT = 8;
+
 
 static inline int
-wrap_bi (int buffer_index)                // wrap buffer index
+wrap_bi (int buffer_index, int mx)                // wrap buffer index
 {
-  return buffer_index & (OUTPUT_RECORD_SIZE - 1);
+  return buffer_index & (mx - 1);
 }
 
 static inline int
-incr_bi (int buffer_index)                // increment buffer index
+incr_bi (int buffer_index, int mx)                // increment buffer index
 {
-  return wrap_bi (buffer_index + 1);
+  return wrap_bi (buffer_index + 1, mx);
 }
 
 static inline int
-decr_bi (int buffer_index)                // decrement buffer index
+decr_bi (int buffer_index, int mx)                // decrement buffer index
 {
-  return wrap_bi (buffer_index - 1);
+  return wrap_bi (buffer_index - 1, mx);
 }
 
 gr_oscope_guts::gr_oscope_guts (double sample_rate, gr_msg_queue_sptr msgq)
@@ -74,8 +86,8 @@ gr_oscope_guts::gr_oscope_guts (double sample_rate, gr_msg_queue_sptr msgq)
     d_buffer[i] = 0;
 
   for (int i = 0; i < MAX_CHANNELS; i++){
-    d_buffer[i] = new float [OUTPUT_RECORD_SIZE];
-    for (int j = 0; j < OUTPUT_RECORD_SIZE; j++)
+    d_buffer[i] = new float [OUTPUT_RECORD_SIZE*SCHART_MULT];
+    for (int j = 0; j < OUTPUT_RECORD_SIZE*SCHART_MULT; j++)
       d_buffer[i][j] = 0.0;
   }
 
@@ -133,18 +145,19 @@ gr_oscope_guts::process_sample (const float *channel_data)
 		assert (0);
 	  }
 
-	  d_obi = incr_bi (d_obi);
+	  d_obi = incr_bi (d_obi, OUTPUT_RECORD_SIZE);
   }
   else
   {
 	  for (int i = 0; i < d_nchannels; i++)
 	  {
-	    for (int j = OUTPUT_RECORD_SIZE-1; j >= 0; j--)
+	    for (int j = (OUTPUT_RECORD_SIZE*SCHART_MULT)-1; j > 0; j--)
 	    {
 			d_buffer[i][j] = d_buffer[i][j-1];
 		}
 		d_buffer[i][0] = channel_data[i];
 	  }
+	  d_trigger_off = 0;
 	  write_output_records();
   }
 }
@@ -183,7 +196,10 @@ gr_oscope_guts::enter_post_trigger ()
 bool
 gr_oscope_guts::found_trigger ()
 {
-  float prev_sample = d_buffer[d_trigger_channel][decr_bi(d_obi)];
+	int mx = d_trigger_mode == gr_TRIG_MODE_STRIPCHART ? OUTPUT_RECORD_SIZE*SCHART_MULT :
+		OUTPUT_RECORD_SIZE;
+		
+  float prev_sample = d_buffer[d_trigger_channel][decr_bi(d_obi, mx)];
   float new_sample = d_buffer[d_trigger_channel][d_obi];
 
   switch (d_trigger_mode){
@@ -224,6 +240,11 @@ gr_oscope_guts::found_trigger ()
 void
 gr_oscope_guts::write_output_records ()
 {
+	int mx;
+	
+	mx = d_trigger_mode == gr_TRIG_MODE_STRIPCHART ? 
+		OUTPUT_RECORD_SIZE*SCHART_MULT : OUTPUT_RECORD_SIZE;
+		
   // if the output queue if full, drop the data like its hot.
   if (d_msgq->full_p())
     return;
@@ -231,17 +252,17 @@ gr_oscope_guts::write_output_records ()
   gr_message_sptr msg = 
     gr_make_message(0,                                         // msg type
             d_nchannels,                                       // arg1 for other side
-            OUTPUT_RECORD_SIZE,                                // arg2 for other side
-            ((d_nchannels * OUTPUT_RECORD_SIZE) + 1) * sizeof(float)); // sizeof payload
+            mx,                                // arg2 for other side
+            ((d_nchannels * mx) + 1) * sizeof(float)); // sizeof payload
 
   float *out = (float *)msg->msg();        // get pointer to raw message buffer
 
   for (int ch = 0; ch < d_nchannels; ch++){
     // note that d_obi + 1 points at the oldest sample in the buffer
-    for (int i = 0; i < OUTPUT_RECORD_SIZE; i++){
-      out[i] = d_buffer[ch][wrap_bi(d_obi + 1 + i)];
+    for (int i = 0; i < mx; i++){
+      out[i] = d_buffer[ch][wrap_bi(d_obi + 1 + i, mx)];
     }
-    out += OUTPUT_RECORD_SIZE;
+    out += mx;
   }
   //Set the last sample as the trigger offset:
   //  The non gl scope sink will not look at this last sample.
@@ -405,5 +426,12 @@ gr_oscope_guts::get_trigger_level () const
 int
 gr_oscope_guts::get_samples_per_output_record () const
 {
-  return OUTPUT_RECORD_SIZE;
+	int mx;
+	
+	mx = OUTPUT_RECORD_SIZE;
+	if (d_trigger_mode == gr_TRIG_MODE_STRIPCHART)
+	{
+		mx = OUTPUT_RECORD_SIZE*SCHART_MULT;
+	}
+  return mx;
 }
diff --git a/gnuradio-core/src/python/gnuradio/gr/qa_fft.py b/gnuradio-core/src/python/gnuradio/gr/qa_fft.py
index 98d80fbb0..e90eb2e7f 100755
--- a/gnuradio-core/src/python/gnuradio/gr/qa_fft.py
+++ b/gnuradio-core/src/python/gnuradio/gr/qa_fft.py
@@ -152,6 +152,60 @@ class test_fft(gr_unittest.TestCase):
         #self.assertComplexTuplesAlmostEqual (expected_result, result_data, 5)
         self.assert_fft_ok2(expected_result, result_data)
 
+    def test_003(self):
+        # Same test as above, only use 2 threads
+
+	tb = gr.top_block()
+        fft_size = 32
+
+        tmp_data = ((4377+4516j),
+                    (-1706.1268310546875+1638.4256591796875j),
+                    (-915.2083740234375+660.69427490234375j),
+                    (-660.370361328125+381.59600830078125j),
+                    (-499.96044921875+238.41630554199219j),
+                    (-462.26748657226562+152.88948059082031j),
+                    (-377.98440551757812+77.5928955078125j),
+                    (-346.85821533203125+47.152004241943359j),
+                    (-295+20j),
+                    (-286.33609008789062-22.257017135620117j),
+                    (-271.52999877929688-33.081821441650391j),
+                    (-224.6358642578125-67.019538879394531j),
+                    (-244.24473571777344-91.524826049804688j),
+                    (-203.09068298339844-108.54627227783203j),
+                    (-198.45195007324219-115.90768432617188j),
+                    (-182.97744750976562-128.12318420410156j),
+                    (-167-180j),
+                    (-130.33688354492188-173.83778381347656j),
+                    (-141.19784545898438-190.28807067871094j),
+                    (-111.09677124023438-214.48896789550781j),
+                    (-70.039543151855469-242.41630554199219j),
+                    (-68.960540771484375-228.30015563964844j),
+                    (-53.049201965332031-291.47097778320312j),
+                    (-28.695289611816406-317.64553833007812j),
+                    (57-300j),
+                    (45.301143646240234-335.69509887695312j),
+                    (91.936195373535156-373.32437133789062j),
+                    (172.09465026855469-439.275146484375j),
+                    (242.24473571777344-504.47515869140625j),
+                    (387.81732177734375-666.6788330078125j),
+                    (689.48553466796875-918.2142333984375j),
+                    (1646.539306640625-1694.1956787109375j))
+
+        src_data = tuple([x/fft_size for x in tmp_data])
+
+        expected_result = tuple([complex(primes[2*i], primes[2*i+1]) for i in range(fft_size)])
+
+        nthreads = 2
+
+        src = gr.vector_source_c(src_data)
+        s2v = gr.stream_to_vector(gr.sizeof_gr_complex, fft_size)
+        fft = gr.fft_vcc(fft_size, False, [], False, nthreads)
+        v2s = gr.vector_to_stream(gr.sizeof_gr_complex, fft_size)
+        dst = gr.vector_sink_c()
+        tb.connect(src, s2v, fft, v2s, dst)
+        tb.run()
+        result_data = dst.data()
+        self.assert_fft_ok2(expected_result, result_data)
 
 if __name__ == '__main__':
     gr_unittest.run(test_fft, "test_fft.xml")
diff --git a/gnuradio-core/src/python/gnuradio/gr/qa_fft_filter.py b/gnuradio-core/src/python/gnuradio/gr/qa_fft_filter.py
index 325495c1d..1e9fdb6a8 100755
--- a/gnuradio-core/src/python/gnuradio/gr/qa_fft_filter.py
+++ b/gnuradio-core/src/python/gnuradio/gr/qa_fft_filter.py
@@ -110,6 +110,23 @@ class test_fft_filter(gr_unittest.TestCase):
         
 
     def test_ccc_002(self):
+        # Test nthreads
+	tb = gr.top_block()
+        src_data = (0,1,2,3,4,5,6,7)
+        taps = (2,)
+        nthreads = 2
+        expected_result = tuple([2 * complex(x) for x in (0,1,2,3,4,5,6,7)])
+        src = gr.vector_source_c(src_data)
+        op = gr.fft_filter_ccc(1, taps, nthreads)
+        dst = gr.vector_sink_c()
+        tb.connect(src, op, dst)
+        tb.run()
+        result_data = dst.data()
+        #print 'expected:', expected_result
+        #print 'results: ', result_data
+        self.assertComplexTuplesAlmostEqual (expected_result, result_data, 5)
+        
+    def test_ccc_003(self):
 	tb = gr.top_block()
         src_data = (0,1,2,3,4,5,6,7)
         taps = (2,)
@@ -124,6 +141,7 @@ class test_fft_filter(gr_unittest.TestCase):
         #print 'results: ', result_data
         self.assertComplexTuplesAlmostEqual (expected_result, result_data, 5)
         
+
     def test_ccc_004(self):
         random.seed(0)
         for i in xrange(25):
@@ -167,6 +185,30 @@ class test_fft_filter(gr_unittest.TestCase):
 
             self.assert_fft_ok2(expected_result, result_data)
 
+    def test_ccc_006(self):
+        # Test decimating with nthreads=2
+        random.seed(0)
+        nthreads = 2
+        for i in xrange(25):
+            # sys.stderr.write("\n>>> Loop = %d\n" % (i,))
+            dec = i + 1
+            src_len = 4*1024
+            src_data = make_random_complex_tuple(src_len)
+            ntaps = int(random.uniform(2, 100))
+            taps = make_random_complex_tuple(ntaps)
+            expected_result = reference_filter_ccc(dec, taps, src_data)
+
+            src = gr.vector_source_c(src_data)
+            op = gr.fft_filter_ccc(dec, taps, nthreads)
+            dst = gr.vector_sink_c()
+            tb = gr.top_block()
+	    tb.connect(src, op, dst)
+            tb.run()
+            del tb
+	    result_data = dst.data()
+
+            self.assert_fft_ok2(expected_result, result_data)
+
     # ----------------------------------------------------------------
     # test _fff version
     # ----------------------------------------------------------------
@@ -202,7 +244,22 @@ class test_fft_filter(gr_unittest.TestCase):
         #print 'results: ', result_data
         self.assertFloatTuplesAlmostEqual (expected_result, result_data, 5)
         
-    def xtest_fff_003(self):
+    def test_fff_003(self):
+        # Test 02 with nthreads
+        tb = gr.top_block()
+        src_data = (0,1,2,3,4,5,6,7)
+        taps = (2,)
+        nthreads = 2
+        expected_result = tuple([2 * float(x) for x in (0,1,2,3,4,5,6,7)])
+        src = gr.vector_source_f(src_data)
+        op = gr.fft_filter_fff(1, taps, nthreads)
+        dst = gr.vector_sink_f()
+        tb.connect(src, op, dst)
+        tb.run()
+        result_data = dst.data()
+        self.assertFloatTuplesAlmostEqual (expected_result, result_data, 5)
+        
+    def xtest_fff_004(self):
         random.seed(0)
         for i in xrange(25):
             sys.stderr.write("\n>>> Loop = %d\n" % (i,))
@@ -232,7 +289,7 @@ class test_fft_filter(gr_unittest.TestCase):
                     actual.write(`x` + '\n')
                 raise
 
-    def xtest_fff_004(self):
+    def xtest_fff_005(self):
         random.seed(0)
         for i in xrange(25):
             sys.stderr.write("\n>>> Loop = %d\n" % (i,))
@@ -252,7 +309,7 @@ class test_fft_filter(gr_unittest.TestCase):
 
             self.assert_fft_float_ok2(expected_result, result_data, abs_eps=2.0)
 
-    def xtest_fff_005(self):
+    def xtest_fff_006(self):
         random.seed(0)
         for i in xrange(25):
             sys.stderr.write("\n>>> Loop = %d\n" % (i,))
@@ -273,6 +330,29 @@ class test_fft_filter(gr_unittest.TestCase):
 
             self.assert_fft_float_ok2(expected_result, result_data)
 
+    def xtest_fff_007(self):
+        # test decimation with nthreads
+        random.seed(0)
+        nthreads = 2
+        for i in xrange(25):
+            sys.stderr.write("\n>>> Loop = %d\n" % (i,))
+            dec = i + 1
+            src_len = 4*1024
+            src_data = make_random_float_tuple(src_len)
+            ntaps = int(random.uniform(2, 100))
+            taps = make_random_float_tuple(ntaps)
+            expected_result = reference_filter_fff(dec, taps, src_data)
+
+            src = gr.vector_source_f(src_data)
+            op = gr.fft_filter_fff(dec, taps, nthreads)
+            dst = gr.vector_sink_f()
+    	    tb = gr.top_block()
+            tb.connect(src, op, dst)
+            tb.run()
+            result_data = dst.data()
+
+            self.assert_fft_float_ok2(expected_result, result_data)
+
     def test_fff_get0(self):
         random.seed(0)
         for i in xrange(25):