moved sgn~ into separate file

initial SSE-support of sgn~

svn path=/trunk/externals/zexy/; revision=4275

4 files changed, 176 insertions, 57 deletions

diff --git a/src/sgn~.c b/src/sgn~.c
new file mode 100644
index 0000000..b390144
--- /dev/null
+++ b/src/sgn~.c
@@ -0,0 +1,173 @@
+/******************************************************
+ *
+ * zexy - implementation file
+ *
+ * copyleft (c) IOhannes m zmölnig
+ *
+ *   2000:forum::für::umläute:2005
+ *
+ *   institute of electronic music and acoustics (iem)
+ *
+ ******************************************************
+ *
+ * license: GNU General Public License v.2
+ *
+ ******************************************************/
+
+/*
+  sgn~: sign of a signal
+  
+  2112:forum::für::umläute:2005
+*/
+
+#include "zexy.h"
+
+typedef struct _sgnTilde
+{
+  t_object x_obj;
+} t_sgnTilde;
+
+
+/* ------------------------ sgn~ ----------------------------- */
+
+static t_class *sgnTilde_class;
+
+static t_int *sgnTilde_perform(t_int *w)
+{
+  t_float *in = (t_float *)(w[1]);
+  t_float *out = (t_float *)(w[2]);
+  int n = (int)(w[3]);
+  t_float x;
+  while (n--) {
+    if ((x=*in++)>0.) *out++=1.;
+    else if	(x<0.) *out++=-1.;
+    else *out++=0.;
+  }
+  
+  return (w+4);
+}
+static t_int *sgnTilde_perform8(t_int *w)
+{
+  t_float *in = (t_float *)(w[1]);
+  t_float *out = (t_float *)(w[2]);
+  int n = (int)(w[3])>>3;
+  t_float x;
+
+    while(n--){
+      /* weirdly enough, the if/else/if/else is a lot faster than ()?:(()?:) */
+      if ((x=in[0])>0.) out[0]=1.; else if(x<0.) out[0]=-1.; else out[0]=0.;
+      if ((x=in[1])>0.) out[1]=1.; else if(x<0.) out[1]=-1.; else out[1]=0.;
+      if ((x=in[2])>0.) out[2]=1.; else if(x<0.) out[2]=-1.; else out[2]=0.;
+      if ((x=in[3])>0.) out[3]=1.; else if(x<0.) out[3]=-1.; else out[3]=0.;
+      if ((x=in[4])>0.) out[4]=1.; else if(x<0.) out[4]=-1.; else out[4]=0.;
+      if ((x=in[5])>0.) out[5]=1.; else if(x<0.) out[5]=-1.; else out[5]=0.;
+      if ((x=in[6])>0.) out[6]=1.; else if(x<0.) out[6]=-1.; else out[6]=0.;
+      if ((x=in[7])>0.) out[7]=1.; else if(x<0.) out[7]=-1.; else out[7]=0.;
+
+      in+=8;
+      out+=8;
+    }
+  
+  return (w+4);
+}
+
+#ifdef __SSE__
+static long l_bitmask[]={0x80000000, 0x80000000, 0x80000000, 0x80000000}; // sign bitmask
+static t_int *sgnTilde_performSSE(t_int *w)
+{
+  __m128 *in = (__m128 *)(w[1]);
+  __m128 *out = (__m128 *)(w[2]);
+
+  __m128 val;
+  int n = (int)(w[3])>>4; // yea, we do 16x loop-unrolling
+
+  const __m128 sgnmask= _mm_loadu_ps((float*)l_bitmask);
+  const __m128 zero   = _mm_setzero_ps();
+  const __m128 one    = _mm_set_ps(1.f, 1.f, 1.f, 1.f);
+
+  __m128 xmm0, xmm1;
+
+  while (n--) {
+
+    val=in[0];
+    xmm0  = _mm_cmpneq_ps(val , zero);// mask for non-zeros
+    xmm1  = _mm_and_ps   (val, sgnmask);// sign (without value)
+    xmm0  = _mm_and_ps   (xmm0, one); // (abs) value: (val==0.f)?0.f:1.f
+    out[0]= _mm_or_ps    (xmm1, xmm0);// merge sign and value
+
+    val=in[1];
+    xmm0  = _mm_cmpneq_ps(val , zero);
+    xmm1  = _mm_and_ps   (val, sgnmask);
+    xmm0  = _mm_and_ps   (xmm0, one);
+    out[1]= _mm_or_ps    (xmm1, xmm0);
+
+    val=in[2];
+    xmm0  = _mm_cmpneq_ps(val , zero);
+    xmm1  = _mm_and_ps   (val, sgnmask);
+    xmm0  = _mm_and_ps   (xmm0, one);
+    out[2]= _mm_or_ps    (xmm1, xmm0);
+
+    val=in[3];
+    xmm0  = _mm_cmpneq_ps(val , zero);
+    xmm1  = _mm_and_ps   (val, sgnmask);
+    xmm0  = _mm_and_ps   (xmm0, one);
+    out[3]= _mm_or_ps    (xmm1, xmm0);
+
+    in +=4;
+    out+=4;
+
+  }
+  return (w+4);
+}
+#endif /* __SSE__ */
+
+static void sgnTilde_dsp(t_sgnTilde *x, t_signal **sp)
+{
+#ifdef __SSE__
+  if(
+     Z_SIMD_CHKBLOCKSIZE(sp[0]->s_n)&&
+     Z_SIMD_CHKALIGN(sp[0]->s_vec)&&
+     Z_SIMD_CHKALIGN(sp[1]->s_vec))
+    {
+      dsp_add(sgnTilde_performSSE, 3, sp[0]->s_vec, sp[1]->s_vec, sp[0]->s_n);
+    } else
+#endif
+    if (sp[0]->s_n & 7) {
+      dsp_add(sgnTilde_perform , 3, sp[0]->s_vec, sp[1]->s_vec, sp[0]->s_n);
+    } else {
+      dsp_add(sgnTilde_perform8, 3, sp[0]->s_vec, sp[1]->s_vec, sp[0]->s_n);
+    }
+
+}
+
+static void sgnTilde_helper(void)
+{
+  post("\n%c sgn~ \t\t:: sign of a signal", HEARTSYMBOL);
+}
+
+static void *sgnTilde_new()
+{
+  t_sgnTilde *x = (t_sgnTilde *)pd_new(sgnTilde_class);
+  outlet_new(&x->x_obj, gensym("signal"));
+  
+  return (x);
+}
+
+static void sgn_tilde_setup(void)
+{
+  sgnTilde_class = class_new(gensym("sgn~"), (t_newmethod)sgnTilde_new, 0,
+			   sizeof(t_sgnTilde), 0, A_DEFFLOAT, 0);
+  class_addmethod(sgnTilde_class, nullfn, gensym("signal"), 0);
+  class_addmethod(sgnTilde_class, (t_method)sgnTilde_dsp, gensym("dsp"), 0);
+  
+  class_addmethod(sgnTilde_class, (t_method)sgnTilde_helper, gensym("help"), 0);
+  class_sethelpsymbol(sgnTilde_class, gensym("zexy/sigbinops+"));
+  zexy_register("sgn~");
+}
+
+
+
+void z_sgn__setup(void)
+{
+  sgn_tilde_setup();
+}
diff --git a/src/z_sigbin.c b/src/z_sigbin.c
index 77de462..50610f7 100644
--- a/src/z_sigbin.c
+++ b/src/z_sigbin.c
@@ -15,68 +15,13 @@
  ******************************************************/
 
 /*
-	finally :: some of the missing binops for signals :: sgn~, >~, <~, ==~, &&~, ||~
+	finally :: some of the missing binops for signals :: >~, <~, ==~, &&~, ||~
 
 	1302:forum::für::umläute:2000
 */
 
 #include "zexy.h"
 
-typedef struct _misc
-{
-  t_object x_obj;
-} t_misc;
-
-
-/* ------------------------ sgn~ ----------------------------- */
-
-static t_class *sigSGN_class;
-
-static t_int *sigSGN_perform(t_int *w)
-{
-  t_float *in = (t_float *)(w[1]);
-  t_float *out = (t_float *)(w[2]);
-  int n = (int)(w[3]);
-  t_float x;
-  
-  while (n--) {
-    if ((x=*in++)>0.) *out++=1.;
-    else if	(x<0.) *out++=-1.;
-    else *out++=0.;
-  }
-  
-  return (w+4);
-}
-
-static void sigSGN_dsp(t_misc *x, t_signal **sp)
-{
-  dsp_add(sigSGN_perform, 3, sp[0]->s_vec, sp[1]->s_vec, sp[0]->s_n);
-}
-
-static void sigSGN_helper(void)
-{
-  post("\n%c sgn~ \t\t:: sign of a signal", HEARTSYMBOL);
-}
-
-static void *sigSGN_new()
-{
-  t_misc *x = (t_misc *)pd_new(sigSGN_class);
-  outlet_new(&x->x_obj, gensym("signal"));
-  
-  return (x);
-}
-
-static void sigSGN_setup(void)
-{
-  sigSGN_class = class_new(gensym("sgn~"), (t_newmethod)sigSGN_new, 0,
-			   sizeof(t_misc), 0, A_DEFFLOAT, 0);
-  class_addmethod(sigSGN_class, nullfn, gensym("signal"), 0);
-  class_addmethod(sigSGN_class, (t_method)sigSGN_dsp, gensym("dsp"), 0);
-  
-  class_addmethod(sigSGN_class, (t_method)sigSGN_helper, gensym("help"), 0);
-  class_sethelpsymbol(sigSGN_class, gensym("zexy/sigbinops+"));
-}
-
 /* ------------------------ relational~ ----------------------------- */
 
 /* ----------------------------- sigGRT ----------------------------- */
@@ -755,7 +700,6 @@ static void sigOR_setup(void)
 
 void z_sigbin_setup(void)
 {
-  sigSGN_setup();
   sigGRT_setup();
   sigLESS_setup();
   sigEQUAL_setup();
diff --git a/src/z_zexy.c b/src/z_zexy.c
index 2eeae8e..f7c1dc3 100644
--- a/src/z_zexy.c
+++ b/src/z_zexy.c
@@ -53,6 +53,7 @@ void z_zexy_setup(void)
 	z_repeat_setup(); /* repeat.c */
 	z_sfplay_setup(); /* sfplay.c */
 	z_sfrecord_setup(); /* sfrecord.c */
+	z_sgn__setup(); /* sgn~.c */
 	z_sigzero__setup(); /* sigzero~.c */
 	z_sort_setup(); /* sort.c */
 	z_step__setup(); /* step~.c */
diff --git a/src/z_zexy.h b/src/z_zexy.h
index 8278e73..6e09612 100644
--- a/src/z_zexy.h
+++ b/src/z_zexy.h
@@ -51,6 +51,7 @@ void z_repack_setup(void); /* repack.c */
 void z_repeat_setup(void); /* repeat.c */
 void z_sfplay_setup(void); /* sfplay.c */
 void z_sfrecord_setup(void); /* sfrecord.c */
+void z_sgn__setup(void); /* sgn~.c */
 void z_sigzero__setup(void); /* sigzero~.c */
 void z_sort_setup(void); /* sort.c */
 void z_step__setup(void); /* step~.c */