Merged eb/gcell -r8215:8243 into trunk. This adds gr-gcell, the GNU

Radio interface to the Cell Broadband Engine.


git-svn-id: http://gnuradio.org/svn/gnuradio/trunk@8244 221aa14e-8319-0410-a670-987f0aec2ac5

4 files changed, 107 insertions, 47 deletions

diff --git a/gcell/src/lib/wrapper/gcp_fft_1d_r2.cc b/gcell/src/lib/wrapper/gcp_fft_1d_r2.cc
index f92ee42c8..07267e303 100644
--- a/gcell/src/lib/wrapper/gcp_fft_1d_r2.cc
+++ b/gcell/src/lib/wrapper/gcp_fft_1d_r2.cc
@@ -30,18 +30,19 @@ static void
 init_jd(gc_job_desc *jd,
 	gc_proc_id_t proc_id,
 	unsigned log2_fft_length,
-	bool forward,
+	bool shift,
 	std::complex<float> *out,
 	const std::complex<float> *in,
-	const std::complex<float> *W)
+	const std::complex<float> *twiddle,
+	const float *window)
 {
   jd->proc_id = proc_id;
   jd->input.nargs = 2;
   jd->output.nargs = 0;
-  jd->eaa.nargs = 3;
+  jd->eaa.nargs = 4;
 
   jd->input.arg[0].u32 = log2_fft_length;
-  jd->input.arg[1].u32 = forward;
+  jd->input.arg[1].u32 = shift;
   unsigned int fft_length = 1 << log2_fft_length;
 
   jd->eaa.arg[0].ea_addr = ptr_to_ea(out);
@@ -52,19 +53,28 @@ init_jd(gc_job_desc *jd,
   jd->eaa.arg[1].direction = GCJD_DMA_GET;
   jd->eaa.arg[1].get_size = sizeof(std::complex<float>) * fft_length;
 
-  jd->eaa.arg[2].ea_addr = ptr_to_ea(const_cast<std::complex<float>*>(W));
+  jd->eaa.arg[2].ea_addr = ptr_to_ea(const_cast<std::complex<float>*>(twiddle));
   jd->eaa.arg[2].direction = GCJD_DMA_GET;
   jd->eaa.arg[2].get_size = sizeof(std::complex<float>) * fft_length / 4;
-}
 
+  jd->eaa.arg[3].ea_addr = ptr_to_ea(const_cast<float*>(window));
+  jd->eaa.arg[3].direction = GCJD_DMA_GET;
+  if (window == 0)
+    jd->eaa.arg[3].get_size = 0;
+  else
+    jd->eaa.arg[3].get_size = sizeof(float) * fft_length;
+}
 
-gc_job_desc *
+  
+gc_job_desc_sptr
 gcp_fft_1d_r2_submit(gc_job_manager_sptr mgr,
 		     unsigned int log2_fft_length,
 		     bool forward,
+		     bool shift,
 		     std::complex<float> *out,
 		     const std::complex<float> *in,
-		     const std::complex<float> *W)
+		     const std::complex<float> *twiddle,
+		     const float *window)
 {
   unsigned int fft_length = 1 << log2_fft_length;
   if (fft_length > 4096)
@@ -74,29 +84,36 @@ gcp_fft_1d_r2_submit(gc_job_manager_sptr mgr,
     throw gc_bad_align("out");
   if ((intptr_t)in & 0xf)
     throw gc_bad_align("in");
-  if ((intptr_t)W & 0xf)
-    throw gc_bad_align("W");
+  if ((intptr_t)twiddle & 0xf)
+    throw gc_bad_align("twiddle");
+  if ((intptr_t)window & 0xf)
+    throw gc_bad_align("window");
+
+  std::string proc_name;
+  if (forward)
+    proc_name = "fwd_fft_1d_r2";
+  else
+    proc_name = "inv_fft_1d_r2";
 
-  gc_proc_id_t fft_id = mgr->lookup_proc("fft_1d_r2");
-  gc_job_desc *jd = mgr->alloc_job_desc();
-  init_jd(jd, fft_id, log2_fft_length, forward, out, in, W);
-  if (!mgr->submit_job(jd)){
+  gc_proc_id_t fft_id = mgr->lookup_proc(proc_name);
+  gc_job_desc_sptr jd = gc_job_manager::alloc_job_desc(mgr);
+  init_jd(jd.get(), fft_id, log2_fft_length, shift, out, in, twiddle, window);
+  if (!mgr->submit_job(jd.get())){
     gc_job_status_t s = jd->status;
-    mgr->free_job_desc(jd);
-    throw gc_bad_submit("fft_1d_r2", s);
+    throw gc_bad_submit(proc_name, s);
   }
   return jd;
 }
 
 void
-gcp_fft_1d_r2_twiddle(unsigned int log2_fft_length, std::complex<float> *W)
+gcp_fft_1d_r2_twiddle(unsigned int log2_fft_length, std::complex<float> *twiddle)
 {
   unsigned int n = 1 << log2_fft_length;
 
-  W[0].real() = 1.0;
-  W[0].imag() = 0.0;
+  twiddle[0].real() = 1.0;
+  twiddle[0].imag() = 0.0;
   for (unsigned i=1; i < n/4; i++){
-    W[i].real() =  cos(i * 2*M_PI/n);
-    W[n/4 - i].imag() = -W[i].real();
+    twiddle[i].real() =  cos(i * 2*M_PI/n);
+    twiddle[n/4 - i].imag() = -twiddle[i].real();
   }
 }
diff --git a/gcell/src/lib/wrapper/gcp_fft_1d_r2.h b/gcell/src/lib/wrapper/gcp_fft_1d_r2.h
index ed1d9e783..1207a5f36 100644
--- a/gcell/src/lib/wrapper/gcp_fft_1d_r2.h
+++ b/gcell/src/lib/wrapper/gcp_fft_1d_r2.h
@@ -25,25 +25,32 @@
 #include <complex>
 
 /*!
- * \brief Submit a job that computes the forward or reverse FFT.
+ * \brief Submit a job that computes the forward or inverse FFT.
  *
  * \param mgr is the job manager instance
  * \param log2_fft_length is the log2 of the fft_length (4 <= x <= 12).
- * \param forward is true to compute the forward xform
+ * \param forward is true to compute the forward transform, else the inverse.
+ * \param shift indicates if an "fftshift" should be applied to the output data
  * \param out is the fft_length output from FFT (must be 16-byte aligned).
  * \param in is the fft_length input to FFT (must be 16-byte aligned).
- * \param W is fft_length/4 twiddle factor input to FFT (must be 16-byte aligned).
+ * \param twiddle is fft_length/4 twiddle factor input to FFT (must be 16-byte aligned).
+ * \param window is the window to be applied to the input data.
+ *    The window length must be either 0 or fft_length (must be 16-byte aligned).
  *
- * Returns a job descriptor which should be passed to wait_job*.
+ * Returns a shared_ptr to a job descriptor which should be passed to wait_job*.
  * Throws an exception in the event of a problem.
+ * This uses the FFTW conventions for scaling.  That is, neither the forward nor inverse
+ * are scaled by 1/fft_length.
  */
-gc_job_desc *
+gc_job_desc_sptr
 gcp_fft_1d_r2_submit(gc_job_manager_sptr mgr,
 		     unsigned int log2_fft_length,
 		     bool forward,
+		     bool shift,
 		     std::complex<float> *out,
 		     const std::complex<float> *in,
-		     const std::complex<float> *W);
+		     const std::complex<float> *twiddle,
+		     const float *window);
 
 /*!
  * \brief Compute twiddle factors 
diff --git a/gcell/src/lib/wrapper/qa_gcp_fft_1d_r2.cc b/gcell/src/lib/wrapper/qa_gcp_fft_1d_r2.cc
index b177edede..404f83657 100644
--- a/gcell/src/lib/wrapper/qa_gcp_fft_1d_r2.cc
+++ b/gcell/src/lib/wrapper/qa_gcp_fft_1d_r2.cc
@@ -80,7 +80,7 @@ qa_gcp_fft_1d_r2::t1()
 #endif
 }
 
-// test reverse FFT
+// test inverse FFT
 void
 qa_gcp_fft_1d_r2::t2()
 {
@@ -101,11 +101,13 @@ qa_gcp_fft_1d_r2::t2()
 void
 qa_gcp_fft_1d_r2::t3()
 {
+  // FIXME Test fwd and inv with windowing option
 }
 
 void
 qa_gcp_fft_1d_r2::t4()
 {
+  // FIXME Test fwd and inv with shift option
 }
 
 static inline float
@@ -178,14 +180,12 @@ qa_gcp_fft_1d_r2::test(gc_job_manager_sptr mgr, int log2_fft_size, bool forward)
 
   // ------------------------------------------------------------------------
   // compute the answer on the cell
-  gc_job_desc *jd = gcp_fft_1d_r2_submit(mgr, log2_fft_size, forward,
-					 cell_out, cell_in, cell_twiddle);
-  if (!mgr->wait_job(jd)){
+  gc_job_desc_sptr jd = gcp_fft_1d_r2_submit(mgr, log2_fft_size, forward, false,
+					     cell_out, cell_in, cell_twiddle, 0);
+  if (!mgr->wait_job(jd.get())){
     fprintf(stderr, "wait_job failed: %s\n", gc_job_status_string(jd->status).c_str());
-    mgr->free_job_desc(jd);
     CPPUNIT_ASSERT(0);
   }
-  mgr->free_job_desc(jd);
 
   // ------------------------------------------------------------------------
   // compute the maximum of the relative error
diff --git a/gcell/src/lib/wrapper/spu/gcs_fft_1d_r2.c b/gcell/src/lib/wrapper/spu/gcs_fft_1d_r2.c
index bf4bdfd20..81e5dfd87 100644
--- a/gcell/src/lib/wrapper/spu/gcs_fft_1d_r2.c
+++ b/gcell/src/lib/wrapper/spu/gcs_fft_1d_r2.c
@@ -21,6 +21,7 @@
 
 #include <gc_declare_proc.h>
 #include <libfft.h>
+#include <assert.h>
 
 /*
  * v is really vector complex<float>
@@ -35,24 +36,59 @@ conjugate_vector(vector float *v, int nelements)
 }
 
 static void
-gcs_fft_1d_r2(const gc_job_direct_args_t *input,
-	      gc_job_direct_args_t *output __attribute__((unused)),
-	      const gc_job_ea_args_t *eaa)
+gcs_fwd_fft_1d_r2(const gc_job_direct_args_t *input,
+		  gc_job_direct_args_t *output __attribute__((unused)),
+		  const gc_job_ea_args_t *eaa)
 {
-  vector float *out = (vector float *) eaa->arg[0].ls_addr;
-  vector float *in = (vector float *) eaa->arg[1].ls_addr;
-  vector float *W = (vector float *) eaa->arg[2].ls_addr;
+  vector float *out = (vector float *) eaa->arg[0].ls_addr;	// complex
+  vector float *in = (vector float *) eaa->arg[1].ls_addr;	// complex
+  vector float *twiddle = (vector float *) eaa->arg[2].ls_addr;	// complex
+  vector float *window = (vector float *) eaa->arg[3].ls_addr;	// float
+
   int log2_fft_length = input->arg[0].u32;
-  int forward = input->arg[1].u32;	// non-zero if forward xform
+  int shift = input->arg[1].u32;	// non-zero if we should apply fftshift
 
-  if (forward){
-    fft_1d_r2(out, in, W, log2_fft_length);
+  if (eaa->arg[3].get_size){	// apply window
+    // FIXME pointwise multiply in *= window
+    assert(0);
   }
-  else {
-    conjugate_vector(in, 1 << (log2_fft_length - 1));
-    fft_1d_r2(out, in, W, log2_fft_length);
-    conjugate_vector(out, 1 << (log2_fft_length - 1));
+
+  fft_1d_r2(out, in, twiddle, log2_fft_length);
+
+  if (shift){
+    // FIXME apply "fftshift" to output data in-place
+    assert(0);
   }
 }
 
-GC_DECLARE_PROC(gcs_fft_1d_r2, "fft_1d_r2");
+GC_DECLARE_PROC(gcs_fwd_fft_1d_r2, "fwd_fft_1d_r2");
+
+static void
+gcs_inv_fft_1d_r2(const gc_job_direct_args_t *input,
+		  gc_job_direct_args_t *output __attribute__((unused)),
+		  const gc_job_ea_args_t *eaa)
+{
+  vector float *out = (vector float *) eaa->arg[0].ls_addr;	// complex
+  vector float *in = (vector float *) eaa->arg[1].ls_addr;	// complex
+  vector float *twiddle = (vector float *) eaa->arg[2].ls_addr;	// complex
+  vector float *window = (vector float *) eaa->arg[3].ls_addr;	// float
+
+  int log2_fft_length = input->arg[0].u32;
+  int shift = input->arg[1].u32;	// non-zero if we should apply fftshift
+
+  if (eaa->arg[3].get_size){	// apply window
+    // FIXME pointwise multiply in *= window
+    assert(0);
+  }
+
+  if (shift){
+    // FIXME apply "fftshift" to input data in-place
+    assert(0);
+  }
+
+  conjugate_vector(in, 1 << (log2_fft_length - 1));
+  fft_1d_r2(out, in, twiddle, log2_fft_length);
+  conjugate_vector(out, 1 << (log2_fft_length - 1));
+}
+
+GC_DECLARE_PROC(gcs_inv_fft_1d_r2, "inv_fft_1d_r2");