Merge multiplications by twiddle with multiplications by i for faster…

… r2c transforms. [empty commit message]
chiajeng · Jan 22, 2007 · 44d6228 · 44d6228
1 parent 37defea
commit 44d6228
Show file tree

Hide file tree

Showing 5 changed files with 68 additions and 0 deletions.
diff --git a/genfft/algsimp.ml b/genfft/algsimp.ml
@@ -115,6 +115,10 @@ end = struct
   and stimesM = function 
     | (Uminus a, b) -> stimesM (a, b) >>= suminusM
     | (a, Uminus b) -> stimesM (a, b) >>= suminusM
+    | (NaN I, CTimes (a, b)) -> stimesM (NaN I, b) >>= 
+	fun ib -> sctimesM (a, ib)
+    | (NaN I, CTimesJ (a, b)) -> stimesM (NaN I, b) >>= 
+	fun ib -> sctimesjM (a, ib)
     | (Num a, Num b) -> snumM (Number.mul a b)
     | (Num a, Times (Num b, c)) -> 
 	snumM (Number.mul a b) >>= fun x -> stimesM (x, c)

diff --git a/genfft/simd.ml b/genfft/simd.ml
@@ -122,8 +122,12 @@ and unparse_expr =
     | Times(NaN CONJ,b) -> op1 "VCONJ" b
     | Times(a,b) ->
 	sprintf "VMUL(%s, %s)" (unparse_expr a) (unparse_expr b)
+    | CTimes(a,Times(NaN I, b)) ->
+	sprintf "VZMULI(%s, %s)" (unparse_expr a) (unparse_expr b)
     | CTimes(a,b) ->
 	sprintf "VZMUL(%s, %s)" (unparse_expr a) (unparse_expr b)
+    | CTimesJ(a,Times(NaN I, b)) ->
+	sprintf "VZMULIJ(%s, %s)" (unparse_expr a) (unparse_expr b)
     | CTimesJ(a,b) ->
 	sprintf "VZMULJ(%s, %s)" (unparse_expr a) (unparse_expr b)
     | Uminus a when !Magic.vneg -> op1 "VNEG" a

diff --git a/simd/simd-altivec.h b/simd/simd-altivec.h
@@ -213,6 +213,30 @@ static inline V VZMULJ(V tx, V sr)
      return VFNMS(ti, si, VMUL(tr, sr));
 }
 
+static inline V VZMULI(V tx, V si)
+{
+     const vector unsigned int real = 
+	  VLIT(0x00010203, 0x00010203, 0x08090a0b, 0x08090a0b);
+     const vector unsigned int imag = 
+	  VLIT(0x04050607, 0x04050607, 0x0c0d0e0f, 0x0c0d0e0f);
+     V sr = VBYI(si);
+     V tr = vec_perm(tx, tx, (vector unsigned char)real);
+     V ti = vec_perm(tx, tx, (vector unsigned char)imag);
+     return VFNMS(ti, si, VMUL(tr, sr));
+}
+
+static inline V VZMULIJ(V tx, V si)
+{
+     const vector unsigned int real = 
+	  VLIT(0x00010203, 0x00010203, 0x08090a0b, 0x08090a0b);
+     const vector unsigned int imag = 
+	  VLIT(0x04050607, 0x04050607, 0x0c0d0e0f, 0x0c0d0e0f);
+     V sr = VBYI(si);
+     V tr = vec_perm(tx, tx, (vector unsigned char)real);
+     V ti = vec_perm(tx, tx, (vector unsigned char)imag);
+     return VFMA(ti, si, VMUL(tr, sr));
+}
+
 /* twiddle storage #1: compact, slower */
 #define VTW1(v,x) \
  {TW_COS, v, x}, {TW_COS, v+1, x}, {TW_SIN, v, x}, {TW_SIN, v+1, x}

diff --git a/simd/simd-sse.h b/simd/simd-sse.h
@@ -232,6 +232,24 @@ static inline V VZMULJ(V tx, V sr)
      return VSUB(tr, VMUL(ti, sr));
 }
 
+static inline V VZMULI(V tx, V sr)
+{
+     V tr = SHUFPS(tx, tx, SHUFVAL(0, 0, 2, 2));
+     V ti = SHUFPS(tx, tx, SHUFVAL(1, 1, 3, 3));
+     ti = VMUL(ti, sr);
+     sr = VBYI(sr);
+     return VSUB(VMUL(tr, sr), ti);
+}
+
+static inline V VZMULIJ(V tx, V sr)
+{
+     V tr = SHUFPS(tx, tx, SHUFVAL(0, 0, 2, 2));
+     V ti = SHUFPS(tx, tx, SHUFVAL(1, 1, 3, 3));
+     ti = VMUL(ti, sr);
+     sr = VBYI(sr);
+     return VADD(VMUL(tr, sr), ti);
+}
+
 #define VFMAI(b, c) VADD(c, VBYI(b))
 #define VFNMSI(b, c) VSUB(c, VBYI(b))
 

diff --git a/simd/simd-sse2.h b/simd/simd-sse2.h
@@ -157,6 +157,24 @@ static inline V VZMULJ(V tx, V sr)
      return VSUB(tr, VMUL(ti, sr));
 }
 
+static inline V VZMULI(V tx, V sr)
+{
+     V tr = UNPCKL(tx, tx);
+     V ti = UNPCKH(tx, tx);
+     ti = VMUL(ti, sr);
+     sr = VBYI(sr);
+     return VSUB(VMUL(tr, sr), ti);
+}
+
+static inline V VZMULIJ(V tx, V sr)
+{
+     V tr = UNPCKL(tx, tx);
+     V ti = UNPCKH(tx, tx);
+     ti = VMUL(ti, sr);
+     sr = VBYI(sr);
+     return VADD(VMUL(tr, sr), ti);
+}
+
 /* twiddle storage #1: compact, slower */
 #define VTW1(v,x) {TW_CEXP, v, x}
 #define TWVL1 1