From b187b797d17ccda6efb853eb93b27aa42474a02a Mon Sep 17 00:00:00 2001
From: Matteo Frigo <athena@fftw.org>
Date: Sun, 12 Feb 2006 15:30:27 -0500
Subject: [PATCH] Added support for t2-style simd split-complex codelets.

---
 dft/simd/Makefile.am          | 4 ++--
 dft/simd/codelets/Makefile.am | 9 +++++++--
 dft/simd/{t1s.c => ts.c}      | 2 +-
 dft/simd/{t1s.h => ts.h}      | 2 +-
 genfft/twiddle.ml             | 8 +++-----
 5 files changed, 14 insertions(+), 11 deletions(-)
 rename dft/simd/{t1s.c => ts.c} (98%)
 rename dft/simd/{t1s.h => ts.h} (96%)

diff --git a/dft/simd/Makefile.am b/dft/simd/Makefile.am
index 27ad0a041..cff784625 100644
--- a/dft/simd/Makefile.am
+++ b/dft/simd/Makefile.am
@@ -2,5 +2,5 @@ AM_CPPFLAGS = -I$(top_srcdir)/kernel -I$(top_srcdir)/dft -I$(top_srcdir)/simd
 noinst_LTLIBRARIES = libdft_simd.la
 SUBDIRS = codelets
 libdft_simd_la_SOURCES = n1b.c n1f.c n2b.c n2f.c n2s.c q1b.c q1f.c t.c	\
-t1s.c n1b.h n1f.h n2b.h n2f.h n2s.h q1b.h q1f.h t1b.h t1f.h t1s.h	\
-t2b.h t2f.h
+ts.c n1b.h n1f.h n2b.h n2f.h n2s.h q1b.h q1f.h t1b.h t1f.h ts.h t2b.h	\
+t2f.h
diff --git a/dft/simd/codelets/Makefile.am b/dft/simd/codelets/Makefile.am
index e0e0aa016..3ac0e9292 100644
--- a/dft/simd/codelets/Makefile.am
+++ b/dft/simd/codelets/Makefile.am
@@ -55,6 +55,7 @@ T2B = t2bv_2.c t2bv_4.c t2bv_8.c t2bv_16.c t2bv_32.c t2bv_64.c
 
 # split-complex codelets
 T1S = t1sv_2.c t1sv_4.c t1sv_8.c t1sv_16.c t1sv_32.c #t1sv_64.c
+T2S = t2sv_2.c t2sv_4.c t2sv_8.c t2sv_16.c t2sv_32.c #t2sv_64.c
 
 ###########################################################################
 # q1fv_<r> is <r> twiddle FFTW_FORWARD FFTs of size <r> (DIF step),
@@ -69,7 +70,7 @@ Q1B = q1bv_2.c q1bv_4.c q1bv_8.c
 
 ###########################################################################
 SIMD_CODELETS = $(N1F) $(N1B) $(N2F) $(N2B) $(N2S) $(T1F) $(T2F)	\
-$(T1B) $(T1S) $(T2B) $(Q1F) $(Q1B)
+$(T1B) $(T2B) $(T1S) $(T2S) $(Q1F) $(Q1B)
 
 if HAVE_SIMD
 ALL_CODELETS = $(SIMD_CODELETS)
@@ -89,6 +90,7 @@ include $(top_srcdir)/support/Makefile.codelets
 
 if MAINTAINER_MODE
 GFLAGS = -simd $(FLAGS_COMMON) -pipeline-latency 8
+FLAGS_T2S=-twiddle-log3 -precompute-twiddles
 
 n1fv_%.c:  $(CODELET_DEPS) $(GEN_NOTW_C)
 	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_NOTW_C) $(GFLAGS) -n $* -name n1fv_$* -include "n1f.h") | $(ADD_DATE) | $(INDENT) >$@
@@ -118,7 +120,10 @@ t2bv_%.c:  $(CODELET_DEPS) $(GEN_TWIDDLE_C)
 	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_TWIDDLE_C) $(GFLAGS) -n $* -name t2bv_$* -include "t2b.h" -sign 1) | $(ADD_DATE) | $(INDENT) >$@
 
 t1sv_%.c:  $(CODELET_DEPS) $(GEN_TWIDDLE)
-	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_TWIDDLE) $(GFLAGS) -n $* -name t1sv_$* -include "t1s.h") | $(ADD_DATE) | $(INDENT) >$@
+	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_TWIDDLE) $(GFLAGS) -n $* -name t1sv_$* -include "ts.h") | $(ADD_DATE) | $(INDENT) >$@
+
+t2sv_%.c:  $(CODELET_DEPS) $(GEN_TWIDDLE)
+	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_TWIDDLE) $(GFLAGS) $(FLAGS_T2S) -n $* -name t2sv_$* -include "ts.h") | $(ADD_DATE) | $(INDENT) >$@
 
 q1fv_%.c:  $(CODELET_DEPS) $(GEN_TWIDSQ_C)
 	($(PRELUDE_COMMANDS_DFT); $(TWOVERS) $(GEN_TWIDSQ_C) $(GFLAGS) -n $* -dif -name q1fv_$* -include "q1f.h") | $(ADD_DATE) | $(INDENT) >$@
diff --git a/dft/simd/t1s.c b/dft/simd/ts.c
similarity index 98%
rename from dft/simd/t1s.c
rename to dft/simd/ts.c
index 1a40ecab6..c020c63de 100644
--- a/dft/simd/t1s.c
+++ b/dft/simd/ts.c
@@ -21,7 +21,7 @@
 #include "codelet-dft.h"
 
 #if HAVE_SIMD
-#include "t1s.h"
+#include "ts.h"
 
 static int okp(const ct_desc *d,
 	       const R *rio, const R *iio, 
diff --git a/dft/simd/t1s.h b/dft/simd/ts.h
similarity index 96%
rename from dft/simd/t1s.h
rename to dft/simd/ts.h
index 1e497348b..6698c5486 100644
--- a/dft/simd/t1s.h
+++ b/dft/simd/ts.h
@@ -28,6 +28,6 @@
 #define VTW VTWS
 #define TWVL TWVLS
 
-#define GENUS X(dft_t1ssimd_genus)
+#define GENUS X(dft_tssimd_genus)
 extern const ct_genus GENUS;
 
diff --git a/genfft/twiddle.ml b/genfft/twiddle.ml
index 2d52435c4..cead7145d 100644
--- a/genfft/twiddle.ml
+++ b/genfft/twiddle.ml
@@ -18,7 +18,7 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  *
  *)
-(* $Id: twiddle.ml,v 1.16 2006-01-05 03:04:27 stevenj Exp $ *)
+(* $Id: twiddle.ml,v 1.17 2006-02-12 20:30:27 athena Exp $ *)
 
 (* policies for loading/computing twiddle factors *)
 open Complex
@@ -38,8 +38,7 @@ type twinstr = (twop * int * int)
 let rec unroll_twfull l = match l with
 | [] -> []
 | (TW_FULL, 0, n) :: b ->
-    (List.flatten
-       (forall [] cons 1 n (fun i -> [(TW_COS, 0, i); (TW_SIN, 0, i)])))
+    (forall [] cons 1 n (fun i -> (TW_CEXP, 0, i)))
     @ unroll_twfull b
 | a :: b -> a :: unroll_twfull b
 
@@ -54,8 +53,7 @@ let twinstr_to_simd_string vl l =
   let one sep = function
     | (TW_NEXT, 1, 0) -> sep ^ "{TW_NEXT, " ^ vl ^ ", 0}"
     | (TW_NEXT, _, _) -> failwith "twinstr_to_simd_string"
-    | (TW_COS, 0, b) -> sep ^ (Printf.sprintf "VTW(%d)" b)
-    | (TW_SIN, 0, b) -> ""
+    | (TW_CEXP, 0, b) -> sep ^ (Printf.sprintf "VTW(%d)" b)
     | _ -> failwith "twinstr_to_simd_string"
   in let rec loop first = function
     | [] -> ""