Implemented buffered direct-r2c, direct-hc2c.

Also, removed some old cruft: * okp() functions were never used and a pain to maintain---now they are gone. * ``m'' in hc2hc and hc2c codelets is now the number of iterations, not the ``logical'' m.
leowu2017 · Dec 20, 2006 · f84c7e6 · f84c7e6
1 parent de904f1
commit f84c7e6
Show file tree

Hide file tree

Showing 18 changed files with 416 additions and 335 deletions.
diff --git a/genfft/gen_hc2c.ml b/genfft/gen_hc2c.ml
@@ -132,10 +132,10 @@ let generate n =
   in
   let body = Block (
     [Decl ("INT", i)],
-    [For (Expr_assign (vi, (CPlus [vm; CUminus (Integer 2)])),
+    [For (Expr_assign (vi, vm),
 	  Binop (" > ", vi, Integer 0),
 	  list_to_comma 
-	    [Expr_assign (vi, CPlus [vi; CUminus (byvl (Integer 2))]);
+	    [Expr_assign (vi, CPlus [vi; CUminus (byvl (Integer 1))]);
 	     Expr_assign (varp, CPlus [varp; byvl vms]);
 	     Expr_assign (vaip, CPlus [vaip; byvl vms]);
 	     Expr_assign (varm, CPlus [varm; CUminus (byvl vms)]);
@@ -166,10 +166,8 @@ let generate n =
       (twinstr_to_string "VL" (twdesc n))
   and desc = 
     Printf.sprintf
-      "static const hc2c_desc desc = {%d, \"%s\", twinstr, &GENUS, %s, %s, %s, %s};\n\n"
+      "static const hc2c_desc desc = {%d, \"%s\", twinstr, &GENUS, %s};\n\n"
       n name (flops_of tree)
-      (stride_to_solverparm !urs) "0"
-      (stride_to_solverparm !ums) 
   and register = "X(khc2c_register)"
 
   in

diff --git a/genfft/gen_hc2hc.ml b/genfft/gen_hc2hc.ml
@@ -57,13 +57,13 @@ let genone sign n transform load store vrs =
   let input = 
     locative_array_c n 
       (C.array_subscript rioarray vrs)
-      (C.array_subscript iioarray (SNeg vrs))
+      (C.array_subscript iioarray vrs)
       locations in
   let output = transform sign n (load n input) in
   let ioloc = 
     locative_array_c n 
       (C.array_subscript rioarray vrs)
-      (C.array_subscript iioarray (SNeg vrs))
+      (C.array_subscript iioarray vrs)
       locations in
   let odag = store n ioloc output in
   let annot = standard_optimizer odag 
@@ -102,11 +102,13 @@ let generate n =
     match !ditdif with
     | DIT -> 
 	genone sign n 
-	  (fun sign n input -> sym2 n (Fft.dft sign n (byw (sym1 n input))))
+	  (fun sign n input -> 
+	     ((sym1 n) @@ (sym2 n)) (Fft.dft sign n (byw input)))
 	  load_array_c store_array_c vrs
     | DIF -> 
 	genone sign n 
-	  (fun sign n input -> sym1 n (byw (Fft.dft sign n (sym2i n input))))
+	  (fun sign n input -> 
+	     byw (Fft.dft sign n (((sym2i n) @@ (sym1 n)) input)))
 	  load_array_c store_array_c vrs
   in
 
@@ -117,10 +119,10 @@ let generate n =
   in
   let body = Block (
     [Decl ("INT", i)],
-    [For (Expr_assign (vi, (CPlus [vm; CUminus (Integer 2)])),
+    [For (Expr_assign (vi, vm),
 	  Binop (" > ", vi, Integer 0),
 	  list_to_comma 
-	    [Expr_assign (vi, CPlus [vi; CUminus (byvl (Integer 2))]);
+	    [Expr_assign (vi, CPlus [vi; CUminus (byvl (Integer 1))]);
 	     Expr_assign (vrioarray, CPlus [vrioarray; byvl vms]);
 	     Expr_assign (viioarray, 
 			  CPlus [viioarray; CUminus (byvl vms)]);
@@ -148,10 +150,8 @@ let generate n =
       (twinstr_to_string "VL" (twdesc n))
   and desc = 
     Printf.sprintf
-      "static const hc2hc_desc desc = {%d, \"%s\", twinstr, &GENUS, %s, %s, %s, %s};\n\n"
+      "static const hc2hc_desc desc = {%d, \"%s\", twinstr, &GENUS, %s};\n\n"
       n name (flops_of tree)
-      (stride_to_solverparm !urs) "0"
-      (stride_to_solverparm !ums) 
   and register = "X(khc2hc_register)"
 
   in

diff --git a/genfft/gen_r2cb.ml b/genfft/gen_r2cb.ml
@@ -145,11 +145,8 @@ let generate n =
 
   in let desc = 
     Printf.sprintf 
-      "static const kr2c_desc desc = { %d, \"%s\", %s, &GENUS, %s, %s, %s, %s, %s };\n\n"
+      "static const kr2c_desc desc = { %d, \"%s\", %s, &GENUS };\n\n"
       n name (flops_of tree) 
-      (stride_to_solverparm !ucsr) (stride_to_solverparm !ucsi)
-      (stride_to_solverparm !urs)
-      "0" "0"
 
   and init =
     (declare_register_fcn name) ^

diff --git a/genfft/gen_r2cf.ml b/genfft/gen_r2cf.ml
@@ -142,11 +142,8 @@ let generate n =
 
   in let desc = 
     Printf.sprintf 
-      "static const kr2c_desc desc = { %d, \"%s\", %s, &GENUS, %s, %s, %s, %s, %s };\n\n"
+      "static const kr2c_desc desc = { %d, \"%s\", %s, &GENUS };\n\n"
       n name (flops_of tree) 
-      (stride_to_solverparm !urs) 
-      (stride_to_solverparm !ucsr) (stride_to_solverparm !ucsi)
-      "0" "0"
 
   and init =
     (declare_register_fcn name) ^

diff --git a/genfft/gen_r2r.ml b/genfft/gen_r2r.ml
@@ -212,7 +212,7 @@ let generate n mode =
 
   in let desc = 
     Printf.sprintf 
-      "static const kr2r_desc desc = { %d, \"%s\", %s, &GENUS, %s, %s, %s, %s, %s };\n\n"
+      "static const kr2r_desc desc = { %d, \"%s\", %s, &GENUS, %s };\n\n"
       n name (flops_of tree) 
       (match mode with
       | RDFT -> "RDFT00"
@@ -227,10 +227,6 @@ let generate n mode =
       | RODFT01 -> "RODFT01"
       | RODFT11 -> "RODFT11"
       | _ -> failwith "must specify a transform kind")
-      (stride_to_solverparm !uistride) 
-      (stride_to_solverparm !uostride)
-      (choose_simd "0" (stride_to_solverparm !uivstride))
-      (choose_simd "0" (stride_to_solverparm !uovstride))
 
   and init =
     (declare_register_fcn name) ^

diff --git a/rdft/codelet-rdft.h b/rdft/codelet-rdft.h
@@ -67,21 +67,15 @@ typedef enum {
 typedef struct kr2c_desc_s kr2c_desc;
 
 typedef struct {
-     int (*okp)(
-	  const kr2c_desc *desc,
-	  const R *R0, const R *R1, const R *Cr, const R *Ci,
-	  INT rs, INT csr, INT csi, INT vl, INT ivs, INT ovs);
      rdft_kind kind;
      INT vl;
 } kr2c_genus;
 
 struct kr2c_desc_s {
-     INT sz;    /* size of transform computed */
+     INT n;    /* size of transform computed */
      const char *nam;
      opcnt ops;
      const kr2c_genus *genus;
-     INT rs, csr, csi;
-     INT ivs, ovs;
 };
 
 typedef void (*kr2c) (R *R0, R *R1, R *Cr, R *Ci,
@@ -93,11 +87,8 @@ void X(kr2c_register)(planner *p, kr2c codelet, const kr2c_desc *desc);
 typedef struct hc2hc_desc_s hc2hc_desc;
 
 typedef struct {
-     int (*okp)(
-	  const struct hc2hc_desc_s *desc,
-	  const R *rio, const R *iio, INT rs, INT vs, INT m, INT ms);
      rdft_kind kind;
-     int vl;
+     INT vl;
 } hc2hc_genus;
 
 struct hc2hc_desc_s {
@@ -106,9 +97,6 @@ struct hc2hc_desc_s {
      const tw_instr *tw;
      const hc2hc_genus *genus;
      opcnt ops;
-     INT rs;
-     INT vs;
-     INT ms;
 };
 
 typedef const R *(*khc2hc) (R *rioarray, R *iioarray, const R *W,
@@ -119,12 +107,8 @@ void X(khc2hc_register)(planner *p, khc2hc codelet, const hc2hc_desc *desc);
 typedef struct hc2c_desc_s hc2c_desc;
 
 typedef struct {
-     int (*okp)(
-	  const struct hc2c_desc_s *desc,
-	  const R *Rp, const R *Ip, const R *Rm, const R *Im, 
-	  INT rs, INT vs, INT m, INT ms);
      rdft_kind kind;
-     int vl;
+     INT vl;
 } hc2c_genus;
 
 struct hc2c_desc_s {
@@ -133,9 +117,6 @@ struct hc2c_desc_s {
      const tw_instr *tw;
      const hc2c_genus *genus;
      opcnt ops;
-     INT rs;
-     INT vs;
-     INT ms;
 };
 
 typedef const R *(*khc2c) (R *Rp, R *Ip, R *Rm, R *Im, const R *W,
@@ -150,22 +131,15 @@ extern const solvtab X(solvtab_rdft_r2cb);
 typedef struct kr2r_desc_s kr2r_desc;
 
 typedef struct {
-     int (*okp)(
-	  const kr2r_desc *desc,
-	  const R *I, const R *O,
-	  INT is, INT os, INT vl, INT ivs, INT ovs);
      INT vl;
 } kr2r_genus;
 
 struct kr2r_desc_s {
-     INT sz;    /* size of transform computed */
+     INT n;    /* size of transform computed */
      const char *nam;
      opcnt ops;
      const kr2r_genus *genus;
      rdft_kind kind;
-     INT is, os;
-     INT ivs;
-     INT ovs;
 };
 
 typedef void (*kr2r) (const R *I, R *O, stride is, stride os,

diff --git a/rdft/codelets/hc2c.c b/rdft/codelets/hc2c.c
@@ -21,21 +21,9 @@
 #include "codelet-rdft.h"
 #include "hc2cf.h"
 
-static int okp(const hc2c_desc *d,
-	       const R *Rp, const R *Ip, const R *Rm, const R *Im, 
-	       INT rs, INT vs, INT m, INT ms)
-{
-     UNUSED(Rp); UNUSED(Ip); UNUSED(Rm); UNUSED(Im); UNUSED(m);
-     return (1
-	     && (!d->rs || (d->rs == rs))
-	     && (!d->vs || (d->vs == vs))
-	     && (!d->ms || (d->ms == ms))
-	  );
-}
-
-const hc2c_genus GENUS = { okp, R2HC, 1 };
+const hc2c_genus GENUS = { R2HC, 1 };
 
 #undef GENUS
 #include "hc2cb.h"
 
-const hc2c_genus GENUS = { okp, HC2R, 1 };
+const hc2c_genus GENUS = { HC2R, 1 };
diff --git a/rdft/codelets/hfb.c b/rdft/codelets/hfb.c
@@ -21,21 +21,9 @@
 #include "codelet-rdft.h"
 #include "hf.h"
 
-static int okp(const hc2hc_desc *d,
-	       const R *rio, const R *iio, 
-	       INT rs, INT vs, INT m, INT ms)
-{
-     UNUSED(rio); UNUSED(iio); UNUSED(m);
-     return (1
-	     && (!d->rs || (d->rs == rs))
-	     && (!d->vs || (d->vs == vs))
-	     && (!d->ms || (d->ms == ms))
-	  );
-}
-
-const hc2hc_genus GENUS = { okp, R2HC, 1 };
+const hc2hc_genus GENUS = { R2HC, 1 };
 
 #undef GENUS
 #include "hb.h"
 
-const hc2hc_genus GENUS = { okp, HC2R, 1 };
+const hc2hc_genus GENUS = { HC2R, 1 };
diff --git a/rdft/codelets/r2c.c b/rdft/codelets/r2c.c
@@ -20,32 +20,18 @@
 
 #include "codelet-rdft.h"
 
-static int okp(const kr2c_desc *d,
-	       const R *R0, const R *R1, const R *Cr, const R *Ci,
-	       INT rs, INT csr, INT csi, INT vl, INT ivs, INT ovs)
-{
-     UNUSED(R0); UNUSED(R1); UNUSED(Cr); UNUSED(Ci); UNUSED(vl);
-     return (1
-	     && (!d->rs || (d->rs == rs))
-	     && (!d->csr || (d->csr == csr))
-	     && (!d->csi || (d->csi == csi))
-	     && (!d->ivs || (d->ivs == ivs))
-	     && (!d->ovs || (d->ovs == ovs))
-	  );
-}
-
 #include "r2cf.h"
-const kr2c_genus GENUS = { okp, R2HC, 1 };
+const kr2c_genus GENUS = { R2HC, 1 };
 #undef GENUS
 
 #include "r2cfII.h"
-const kr2c_genus GENUS = { okp, R2HCII, 1 };
+const kr2c_genus GENUS = { R2HCII, 1 };
 #undef GENUS
 
 #include "r2cb.h"
-const kr2c_genus GENUS = { okp, HC2R, 1 };
+const kr2c_genus GENUS = { HC2R, 1 };
 #undef GENUS
 
 #include "r2cbIII.h"
-const kr2c_genus GENUS = { okp, HC2RIII, 1 };
+const kr2c_genus GENUS = { HC2RIII, 1 };
 #undef GENUS
diff --git a/rdft/codelets/r2r.c b/rdft/codelets/r2r.c
@@ -21,18 +21,4 @@
 #include "codelet-rdft.h"
 #include "r2r.h"
 
-static int okp(const kr2r_desc *d,
-	       const R *I,
-	       const R *O,
-	       INT is, INT os, INT vl, INT ivs, INT ovs)
-{
-     UNUSED(I); UNUSED(O); UNUSED(vl);
-     return (1
-	     && (!d->is || (d->is == is))
-	     && (!d->os || (d->os == os))
-	     && (!d->ivs || (d->ivs == ivs))
-	     && (!d->ovs || (d->ovs == ovs))
-	  );
-}
-
-const kr2r_genus GENUS = { okp, 1 };
+const kr2r_genus GENUS = { 1 };