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

10 files changed, 273 insertions, 75 deletions

diff --git a/gcell/src/lib/runtime/Makefile.am b/gcell/src/lib/runtime/Makefile.am
index a68d2bcd0..89d6f1bb9 100644
--- a/gcell/src/lib/runtime/Makefile.am
+++ b/gcell/src/lib/runtime/Makefile.am
@@ -32,6 +32,7 @@ dist_bin_SCRIPTS = gcell-embedspu-libtool
 noinst_LTLIBRARIES = libruntime.la libruntime-qa.la
 
 libruntime_la_SOURCES = \
+	gc_aligned_alloc.cc \
 	gc_job_manager.cc \
 	gc_job_manager_impl.cc \
 	gc_jd_queue.c \
@@ -46,6 +47,7 @@ libruntime_qa_la_SOURCES = \
 
 
 gcellinclude_HEADERS = \
+	gc_aligned_alloc.h \
 	gc_job_manager.h
 
 noinst_HEADERS = \
diff --git a/gcell/src/lib/runtime/gc_aligned_alloc.cc b/gcell/src/lib/runtime/gc_aligned_alloc.cc
new file mode 100644
index 000000000..fa20a6443
--- /dev/null
+++ b/gcell/src/lib/runtime/gc_aligned_alloc.cc
@@ -0,0 +1,54 @@
+/* -*- c++ -*- */
+/*
+ * Copyright 2008 Free Software Foundation, Inc.
+ * 
+ * This file is part of GNU Radio
+ * 
+ * GNU Radio is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3, or (at your option)
+ * any later version.
+ * 
+ * GNU Radio is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ * 
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+#include <gc_aligned_alloc.h>
+#include <stdlib.h>
+#include <stdexcept>
+
+// custom deleter of anything that can be freed with "free"
+class free_deleter {
+public:
+  void operator()(void *p) {
+    free(p);
+  }
+};
+
+void *
+gc_aligned_alloc(size_t size, size_t alignment)
+{
+  void *p = 0;
+  if (posix_memalign(&p, alignment, size) != 0){
+    perror("posix_memalign");
+    throw std::runtime_error("memory");
+  }
+  memset(p, 0, size);		// zero the memory
+  return p;
+}
+
+boost::shared_ptr<void>
+gc_aligned_alloc_sptr(size_t size, size_t alignment)
+{
+  return boost::shared_ptr<void>(gc_aligned_alloc(size, alignment),
+				 free_deleter());
+}
diff --git a/gcell/src/lib/runtime/gc_aligned_alloc.h b/gcell/src/lib/runtime/gc_aligned_alloc.h
new file mode 100644
index 000000000..bdc21c278
--- /dev/null
+++ b/gcell/src/lib/runtime/gc_aligned_alloc.h
@@ -0,0 +1,52 @@
+/* -*- c++ -*- */
+/*
+ * Copyright 2008 Free Software Foundation, Inc.
+ * 
+ * This file is part of GNU Radio
+ * 
+ * GNU Radio is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3, or (at your option)
+ * any later version.
+ * 
+ * GNU Radio is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ * 
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+#ifndef INCLUDED_GC_ALIGNED_ALLOC_H
+#define INCLUDED_GC_ALIGNED_ALLOC_H
+
+#include <boost/shared_ptr.hpp>
+
+/*!
+ * \brief Return pointer to chunk of storage of size size bytes.
+ * The allocation will be aligned to an \p alignment boundary.
+ *
+ * \param size is the number of bytes to allocate
+ * \param alignment is the minimum storage alignment in bytes; must be a power of 2.
+ *
+ * Throws if can't allocate memory.  The storage should be freed
+ * with "free" when done.  The memory is initialized to zero.
+ */
+void *
+gc_aligned_alloc(size_t size, size_t alignment = 128);
+
+/*!
+ * \brief Return boost::shared_ptr to chunk of storage of size size bytes.
+ * The allocation will be aligned to an \p alignment boundary.
+ *
+ * \param size is the number of bytes to allocate
+ * \param alignment is the minimum storage alignment in bytes; must be a power of 2.
+ *
+ * Throws if can't allocate memory.  The storage should be freed
+ * with "free" when done.  The memory is initialized to zero.
+ */
+boost::shared_ptr<void>
+gc_aligned_alloc_sptr(size_t size, size_t alignment = 128);
+
+#endif /* INCLUDED_GC_ALIGNED_ALLOC_H */
diff --git a/gcell/src/lib/runtime/gc_job_manager.cc b/gcell/src/lib/runtime/gc_job_manager.cc
index 9ede5e156..ac2e989a4 100644
--- a/gcell/src/lib/runtime/gc_job_manager.cc
+++ b/gcell/src/lib/runtime/gc_job_manager.cc
@@ -31,6 +31,19 @@
 static boost::weak_ptr<gc_job_manager> s_singleton;
 
 
+// custom deleter of gc_job_desc allocated via alloc_job_desc_sptr
+class job_desc_deleter {
+  gc_job_manager_sptr	d_mgr;
+public:
+  job_desc_deleter(gc_job_manager_sptr mgr) : d_mgr(mgr) {}
+
+  void operator()(gc_job_desc *jd) {
+    d_mgr->free_job_desc(jd);
+  }
+};
+
+
+
 gc_job_manager_sptr
 gc_make_job_manager(const gc_jm_options *options)
 {
@@ -71,6 +84,19 @@ gc_job_manager::singleton()
   return gc_job_manager_sptr(s_singleton);
 }
 
+gc_job_desc_sptr 
+gc_job_manager::make_jd_sptr(gc_job_manager_sptr mgr, gc_job_desc *jd)
+{
+  return gc_job_desc_sptr(jd, job_desc_deleter(mgr));
+}
+
+gc_job_desc_sptr 
+gc_job_manager::alloc_job_desc(gc_job_manager_sptr mgr)
+{
+  return make_jd_sptr(mgr, mgr->alloc_job_desc());
+}
+
+
 // ------------------------------------------------------------------------
 
 
diff --git a/gcell/src/lib/runtime/gc_job_manager.h b/gcell/src/lib/runtime/gc_job_manager.h
index aa30dc24b..67abce7ed 100644
--- a/gcell/src/lib/runtime/gc_job_manager.h
+++ b/gcell/src/lib/runtime/gc_job_manager.h
@@ -33,6 +33,7 @@
 class gc_job_manager;
 typedef boost::shared_ptr<gc_job_manager> gc_job_manager_sptr;
 typedef boost::shared_ptr<spe_program_handle_t> spe_program_handle_sptr;
+typedef boost::shared_ptr<gc_job_desc> gc_job_desc_sptr;
 
 /*!
  * \brief Return a boost::shared_ptr to an spe_program_handle_t
@@ -86,10 +87,19 @@ struct gc_jm_options {
 
   gc_jm_options() :
     max_jobs(0), max_client_threads(0), nspes(0),
-    gang_schedule(true), use_affinity(false),
+    gang_schedule(false), use_affinity(false),
     enable_logging(false), log2_nlog_entries(12)
   {
   }
+
+  gc_jm_options(spe_program_handle_sptr program_handle_,
+		unsigned int nspes_ = 0) :
+    max_jobs(0), max_client_threads(0), nspes(nspes_),
+    gang_schedule(false), use_affinity(false),
+    enable_logging(false), log2_nlog_entries(12),
+    program_handle(program_handle_)
+  {
+  }
 };
 
 enum gc_wait_mode {
@@ -236,6 +246,11 @@ public:
    */
   virtual std::vector<std::string> proc_names() = 0;
 
+  virtual void set_debug(int debug);
+  virtual int debug();
+
+  /* ----- static methods ----- */
+
   /*!
    * \brief Set the singleton gc_job_manager instance.
    * \param mgr is the job manager instance.
@@ -256,9 +271,15 @@ public:
    */
   static gc_job_manager_sptr singleton();
 
+  /*!
+   * \brief return a boost::shared_ptr to a job descriptor.
+   */
+  static gc_job_desc_sptr make_jd_sptr(gc_job_manager_sptr mgr, gc_job_desc *jd);
 
-  virtual void set_debug(int debug);
-  virtual int debug();
+  /*!
+   * \brief allocate a job descriptor and return a boost::shared_ptr to it.
+   */
+  static gc_job_desc_sptr alloc_job_desc(gc_job_manager_sptr mgr);
 };
 
 
diff --git a/gcell/src/lib/runtime/gc_job_manager_impl.cc b/gcell/src/lib/runtime/gc_job_manager_impl.cc
index 59deb4ae5..9c859511b 100644
--- a/gcell/src/lib/runtime/gc_job_manager_impl.cc
+++ b/gcell/src/lib/runtime/gc_job_manager_impl.cc
@@ -25,7 +25,7 @@
 #include <gc_job_manager_impl.h>
 #include <gc_mbox.h>
 #include <gc_proc_def_utils.h>
-
+#include <gc_aligned_alloc.h>
 #include <stdio.h>
 #include <stdexcept>
 #include <stdlib.h>
@@ -85,23 +85,6 @@ client_key_destructor(void *p)
   ((gc_client_thread_info *) p)->d_free = 1;
 }
 
-/*
- * Return pointer to cache-aligned chunk of storage of size size bytes.
- * Throw if can't allocate memory.  The storage should be freed
- * with "free" when done.  The memory is initialized to zero.
- */
-static void *
-aligned_alloc(size_t size, size_t alignment = CACHE_LINE_SIZE)
-{
-  void *p = 0;
-  if (posix_memalign(&p, alignment, size) != 0){
-    perror("posix_memalign");
-    throw std::runtime_error("memory");
-  }
-  memset(p, 0, size);		// zero the memory
-  return p;
-}
-
 static bool
 is_power_of_2(uint32_t x)
 {
@@ -196,7 +179,7 @@ gc_job_manager_impl::gc_job_manager_impl(const gc_jm_options *options)
   // ----------------------------------------------------------------
   // initalize the job queue
   
-  d_queue = (gc_jd_queue_t *) aligned_alloc(sizeof(gc_jd_queue_t));
+  d_queue = (gc_jd_queue_t *) gc_aligned_alloc(sizeof(gc_jd_queue_t), CACHE_LINE_SIZE);
   _d_queue_boost =
     boost::shared_ptr<void>((void *) d_queue, free_deleter());
   gc_jd_queue_init(d_queue);
@@ -208,15 +191,15 @@ gc_job_manager_impl::gc_job_manager_impl(const gc_jm_options *options)
   // 1 spu_arg struct for each SPE
   assert(sizeof(gc_spu_args_t) % 16 == 0);
   d_spu_args =
-    (gc_spu_args_t *) aligned_alloc(MAX_SPES * sizeof(gc_spu_args_t), 16);
+    (gc_spu_args_t *) gc_aligned_alloc(MAX_SPES * sizeof(gc_spu_args_t), 16);
   _d_spu_args_boost =
     boost::shared_ptr<void>((void *) d_spu_args, free_deleter());
 
   // 2 completion info structs for each SPE (we double buffer them)
   assert(sizeof(gc_comp_info_t) % CACHE_LINE_SIZE == 0);
   d_comp_info =
-    (gc_comp_info_t *) aligned_alloc(2 * MAX_SPES * sizeof(gc_comp_info_t),
-				     CACHE_LINE_SIZE);
+    (gc_comp_info_t *) gc_aligned_alloc(2 * MAX_SPES * sizeof(gc_comp_info_t),
+					CACHE_LINE_SIZE);
   _d_comp_info_boost =
     boost::shared_ptr<void>((void *) d_comp_info, free_deleter());
 
@@ -269,7 +252,7 @@ gc_job_manager_impl::gc_job_manager_impl(const gc_jm_options *options)
   // ----------------------------------------------------------------
   // initalize the free list of job descriptors
   
-  d_free_list = (gc_jd_stack_t *) aligned_alloc(sizeof(gc_jd_stack_t));
+  d_free_list = (gc_jd_stack_t *) gc_aligned_alloc(sizeof(gc_jd_stack_t), CACHE_LINE_SIZE);
   // This ensures that the memory associated with d_free_list is
   // automatically freed in the destructor or if an exception occurs
   // here in the constructor.
@@ -283,7 +266,7 @@ gc_job_manager_impl::gc_job_manager_impl(const gc_jm_options *options)
   }
 
   // Initialize the array of job descriptors.
-  d_jd = (gc_job_desc_t *) aligned_alloc(sizeof(d_jd[0]) * d_options.max_jobs);
+  d_jd = (gc_job_desc_t *) gc_aligned_alloc(sizeof(d_jd[0]) * d_options.max_jobs, CACHE_LINE_SIZE);
   _d_jd_boost = boost::shared_ptr<void>((void *) d_jd, free_deleter());
 
 
@@ -317,7 +300,7 @@ gc_job_manager_impl::gc_job_manager_impl(const gc_jm_options *options)
 
   // allocate all bitvectors in a single cache-aligned chunk
   size_t nlongs = d_bvlen * d_options.max_client_threads;
-  void *p = aligned_alloc(nlongs * sizeof(unsigned long));
+  void *p = gc_aligned_alloc(nlongs * sizeof(unsigned long), CACHE_LINE_SIZE);
   _d_all_bitvectors = boost::shared_ptr<void>(p, free_deleter());
 
   // Now point the gc_client_thread_info bitvectors into this storage
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");