From 2cecd0d85296320f8e254fc8046cb8ff24bdcfe4 Mon Sep 17 00:00:00 2001
From: Ed J <mohawk2@users.noreply.github.com>
Date: Thu, 28 Nov 2024 20:45:26 +0000
Subject: [PATCH] use fftw_malloc not _alloc_real, broadcastloop c2r so only
 malloc once

---
 Makefile.PL |  4 ++--
 fftw3.pd    | 20 ++++++++++----------
 2 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/Makefile.PL b/Makefile.PL
index 162fc71..a51e281 100644
--- a/Makefile.PL
+++ b/Makefile.PL
@@ -54,10 +54,10 @@ $descriptor{INC} = '' unless defined $descriptor{INC};
 $descriptor{INC} .= " $cflags";
 
 $descriptor{PREREQ_PM} = {
-  'PDL' => '2.049', # as_native
+  'PDL' => '2.075', # broadcastloop
 };
 $descriptor{CONFIGURE_REQUIRES} = {
-  'PDL' => '2.049',
+  'PDL' => '2.075',
   'IPC::Run'       =>0,
 #  'Alien::FFTW3'   =>0,
 };
diff --git a/fftw3.pd b/fftw3.pd
index 80cdf54..ca7b74a 100644
--- a/fftw3.pd
+++ b/fftw3.pd
@@ -52,22 +52,22 @@ static_assert_fftw(sizeof($GENERIC()) == sizeof($TGC(fftwf_,fftw_)complex));
 $TGC(fftwf_,fftw_)plan plan = INT2PTR($TGC(fftwf_,fftw_)plan, $COMP(plan));
 // FFTW inverse real transforms clobber their input. I thus make a new
 // buffer and transform from there
-unsigned long nelem = 1;
-PDL_Indx i, rank = $PRIV(vtable)->par_realdims[0];
-for( i=0; i<rank; i++ )
-  nelem *= $PDL(complexv)->dims[i];
-unsigned long elem_scale = sizeof($GENERIC()) / sizeof( $TGC(float,double) ); /* native complex */
-void *input_copy = $TGC(fftwf_,fftw_)alloc_real( nelem * elem_scale );
-memcpy( input_copy, $P(complexv), sizeof($GENERIC()) * nelem );
-$TGC(fftwf_,fftw_)execute_dft_c2r( plan, (void*)input_copy, (void*)$P(real) );
-$TGC(fftwf_,fftw_)free( input_copy );
+PDL_Indx i, nbytes = sizeof($GENERIC());
+PDL_Indx rank = $PRIV(vtable)->par_realdims[0], *dims = $PDL(complexv)->dims;
+for (i=0; i<rank; i++) nbytes *= dims[i];
+void *input_copy = fftw_malloc(nbytes);
+broadcastloop %{
+  memcpy(input_copy, $P(complexv), nbytes);
+  $TGC(fftwf_,fftw_)execute_dft_c2r(plan, input_copy, (void*)$P(real));
+%}
+fftw_free(input_copy);
 EOF
 my $TEMPLATE_COMPLEX = <<'EOF';
 // This is the template used by PP to generate the FFTW routines.
 // make sure the PDL data type I'm using matches the FFTW data type
 static_assert_fftw(sizeof($GENERIC())*2 == sizeof($TFD(fftwf_,fftw_)complex));
 $TFD(fftwf_,fftw_)plan plan = INT2PTR($TFD(fftwf_,fftw_)plan, $COMP(plan));
-$TFD(fftwf_,fftw_)execute_dft( plan, (void*)$P(in), (void*)$P(out) );
+$TFD(fftwf_,fftw_)execute_dft(plan, (void*)$P(in), (void*)$P(out));
 EOF
 
 # I define up to rank-10 FFTs. This is annoyingly arbitrary, but hopefully