[Commit-gnuradio] r8968 - in gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib: filter general

[eb]

Author: eb
Date: 2008-07-21 21:43:45 -0600 (Mon, 21 Jul 2008)
New Revision: 8968

Added:
   gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.c
   gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.h
Modified:
   gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/Makefile.am
   
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc
   gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/general/gr_math.h
Log:
work-in-progress on altivec

Modified: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/Makefile.am
===================================================================
--- 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/Makefile.am    
    2008-07-22 00:21:02 UTC (rev 8967)
+++ 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/Makefile.am    
    2008-07-22 03:43:45 UTC (rev 8968)
@@ -170,7 +170,8 @@
        sysconfig_powerpc.cc \
        gr_fir_sysconfig_powerpc.cc \
        gr_cpu_powerpc.cc \
-       gr_fir_fff_vmx.cc
+       gr_fir_fff_vmx.cc \
+       gr_altivec.c
 
 powerpc_qa_CODE = \
        qa_dotprod_powerpc.cc
@@ -262,6 +263,7 @@
        float_dotprod_generic.h         \
        float_dotprod_x86.h             \
        gr_adaptive_fir_ccf.h           \
+       gr_altivec.h                    \
        gr_cma_equalizer_cc.h           \
        gr_cpu.h                        \
        gr_fft_filter_ccc.h             \

Added: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.c
===================================================================
--- 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.c   
                            (rev 0)
+++ 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.c   
    2008-07-22 03:43:45 UTC (rev 8968)
@@ -0,0 +1,38 @@
+/* -*- c++ -*- */
+/*
+ * Copyright 2008 Free Software Foundation, Inc.
+ * 
+ * This file is part of GNU Radio
+ * 
+ * GNU Radio is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3, or (at your option)
+ * any later version.
+ * 
+ * GNU Radio is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ * 
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#include <gr_altivec.h>
+
+void
+gr_print_vector_float(FILE *fp, vector float v)
+{
+  union v_float_u      u;
+  u.v = v;
+  fprintf(fp, "{ %f, %f, %f, %f }", u.f[0], u.f[1], u.f[2], u.f[3]);
+}
+  
+void
+gr_pvf(FILE *fp, const char *label, vector float v)
+{
+  fprintf(fp, "%s = ", label);
+  gr_print_vector_float(fp, v);
+  putc('\n', fp);
+}


Property changes on: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.c
___________________________________________________________________
Name: svn:eol-style
   + native

Added: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.h
===================================================================
--- 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.h   
                            (rev 0)
+++ 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.h   
    2008-07-22 03:43:45 UTC (rev 8968)
@@ -0,0 +1,57 @@
+/* -*- c++ -*- */
+/*
+ * Copyright 2008 Free Software Foundation, Inc.
+ * 
+ * This file is part of GNU Radio
+ * 
+ * GNU Radio is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3, or (at your option)
+ * any later version.
+ * 
+ * GNU Radio is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ * 
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+#ifndef INCLUDED_GR_ALTIVEC_H
+#define INCLUDED_GR_ALTIVEC_H
+
+#include <altivec.h>
+#include <stddef.h>
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define VS             sizeof(vector float)
+#define FLOATS_PER_VEC (sizeof(vector float)/sizeof(float))
+
+union v_float_u {
+  vector float v;
+  float                f[FLOATS_PER_VEC];
+};
+
+void gr_print_vector_float(FILE *fp, vector float v);
+void gr_pvf(FILE *fp, const char *label, vector float v);
+
+static inline float
+horizontal_add_f(vector float v)
+{
+  union v_float_u u;
+  vector float   t0 = vec_add(v, vec_sld(v, v, 8));
+  vector float   t1 = vec_add(t0, vec_sld(t0, t0, 4));
+  u.v = t1;
+  return u.f[0];
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* INCLUDED_GR_ALTIVEC_H */


Property changes on: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_altivec.h
___________________________________________________________________
Name: svn:eol-style
   + native

Modified: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc
===================================================================
--- 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc
  2008-07-22 00:21:02 UTC (rev 8967)
+++ 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/filter/gr_fir_fff_vmx.cc
  2008-07-22 03:43:45 UTC (rev 8968)
@@ -23,76 +23,14 @@
 #include <config.h>
 #endif
 #include <gr_fir_fff_vmx.h>
-#include <altivec.h>
 #include <stdlib.h>
 #include <stdexcept>
 #include <assert.h>
+#include <gr_math.h>
+#include <gr_altivec.h>
 
+extern "C" {
 
-static const size_t VS = sizeof(vector float);
-static const size_t FLOATS_PER_VEC = 4;
-
-union v_float_u {
-  vector float v;
-  float                f[FLOATS_PER_VEC];
-};
-
-void
-print_vector_float(FILE *fp, vector float v)
-{
-  v_float_u    u;
-  u.v = v;
-  fprintf(fp, "{ %f, %f, %f, %f }", u.f[0], u.f[1], u.f[2], u.f[3]);
-}
-  
-void
-print_vector_float_nl(FILE *fp, vector float v)
-{
-  print_vector_float(fp, v);
-  putc('\n', fp);
-}
-
-void
-pvf(FILE *fp, const char *name, vector float v)
-{
-  fprintf(fp, "%s = ", name);
-  print_vector_float_nl(fp, v);
-}
-
-static float
-horizontal_add_f(vector float v)
-{
-  v_float_u    u;
-  vector float t0 = vec_add(v, vec_sld(v, v, 8));
-  vector float t1 = vec_add(t0, vec_sld(t0, t0, 4));
-  u.v = t1;
-  return u.f[0];
-}
-
-static inline size_t
-gr_round_down(size_t x, size_t pow2)
-{
-  return x & -pow2;
-}
-
-static inline size_t
-gr_round_up(size_t x, size_t pow2)
-{
-  return gr_round_down(x + pow2 - 1, pow2);
-}
-
-static inline size_t
-gr_modulo(size_t x, size_t pow2)
-{
-  return x & (pow2 - 1);
-}
-
-static inline size_t
-gr_modulo_neg(size_t x, size_t pow2)
-{
-  return pow2 - gr_modulo(x, pow2);
-}
-
 #if 0
 
 float
@@ -105,58 +43,8 @@
   return sum;
 }
 
-#elif 0
-/*
- *  preconditions:
- *
- *    n > 0 and a multiple of 4
- *    a  4-byte aligned
- *    b  4-byte aligned
- */
-float
-dotprod_fff_vmx(const float *a, const float *b, size_t n)
-{
-  static const size_t FLOATS_PER_LOOP = 1 * FLOATS_PER_VEC;
-
-  size_t loop_cnt = n / FLOATS_PER_LOOP;
-  vector float acc0 = {0, 0, 0, 0};
-
-  vector unsigned char lvsl_a;
-  vector float msq_a0; // most significant quadword
-  vector float lsq_a0; // least significant quadword
-
-  vector unsigned char lvsl_b;
-  vector float msq_b0; // most significant quadword
-  vector float lsq_b0; // least significant quadword
-
-  lvsl_a = vec_lvsl(0, a);
-  msq_a0 = vec_ld(0, a);
-  a += FLOATS_PER_VEC;
-
-  lvsl_b = vec_lvsl(0, b);
-  msq_b0 = vec_ld(0, b);
-  b += FLOATS_PER_VEC;
-  
-  for (size_t i = 0; i < loop_cnt; i++){
-    lsq_a0 = vec_ld(0, a);
-    lsq_b0 = vec_ld(0, b);
-    a += FLOATS_PER_VEC;
-    b += FLOATS_PER_VEC;
-
-    vector float va = vec_perm(msq_a0, lsq_a0, lvsl_a);
-    msq_a0 = lsq_a0;
-
-    vector float vb = vec_perm(msq_b0, lsq_b0, lvsl_b);
-    msq_b0 = lsq_b0;
-
-    acc0 = vec_madd(va, vb, acc0);
-  }
-
-  return horizontal_add_f(acc0);
-}
-
 #else
-/*
+/*
  *  preconditions:
  *
  *    n > 0 and a multiple of 4
@@ -171,6 +59,7 @@
 
   static const size_t UNROLL_CNT = 4;
 
+  n = gr_p2_round_down(n, 4);
   size_t loop_cnt = n / (UNROLL_CNT * FLOATS_PER_VEC);
   size_t nleft = n % (UNROLL_CNT * FLOATS_PER_VEC);
 
@@ -189,71 +78,75 @@
 
   // wind in
 
-  p0 = vec_ld(0*VS, a);
-  p1 = vec_ld(1*VS, a);
-  p2 = vec_ld(2*VS, a);
-  p3 = vec_ld(3*VS, a);
+  register int r0vs = 0 * VS;
+  register int r1vs = 1 * VS;
+  register int r2vs = 2 * VS;
+  register int r3vs = 3 * VS;
 
+  p0 = vec_ld(r0vs, a);
+  p1 = vec_ld(r1vs, a);
+  p2 = vec_ld(r2vs, a);
+  p3 = vec_ld(r3vs, a);
+  a += UNROLL_CNT;
+
   a0 = vec_perm(p0, p1, lvsl_a);
-  b0 = vec_ld(0*VS, b);
-  p0 = vec_ld((UNROLL_CNT + 0)*VS, a);
+  b0 = vec_ld(r0vs, b);
+  p0 = vec_ld(r0vs, a);
 
   for (size_t i = 0; i < loop_cnt; i++){
 
     a1 = vec_perm(p1, p2, lvsl_a);
-    b1 = vec_ld(1*VS, b);
-    p1 = vec_ld((UNROLL_CNT + 1)*VS, a);
+    b1 = vec_ld(r1vs, b);
+    p1 = vec_ld(r1vs, a);
     acc0 = vec_madd(a0, b0, acc0);
 
     a2 = vec_perm(p2, p3, lvsl_a);
-    b2 = vec_ld(2*VS, b);
-    p2 = vec_ld((UNROLL_CNT + 2)*VS, a);
+    b2 = vec_ld(r2vs, b);
+    p2 = vec_ld(r2vs, a);
     acc1 = vec_madd(a1, b1, acc1);
 
     a3 = vec_perm(p3, p0, lvsl_a);
-    b3 = vec_ld(3*VS, b);
-    p3 = vec_ld((UNROLL_CNT + 3)*VS, a);
+    b3 = vec_ld(r3vs, b);
+    p3 = vec_ld(r3vs, a);
     acc2 = vec_madd(a2, b2, acc2);
 
     a += UNROLL_CNT;
     b += UNROLL_CNT;
 
     a0 = vec_perm(p0, p1, lvsl_a);
-    b0 = vec_ld(0*VS, b);
-    p0 = vec_ld((UNROLL_CNT + 0)*VS, a);
+    b0 = vec_ld(r0vs, b);
+    p0 = vec_ld(r0vs, a);
     acc3 = vec_madd(a3, b3, acc3);
   }
 
-  assert((nleft % 4) == 0);
-
-  switch (nleft/4){
+  /*
+   * The compiler ought to be able to figure out that 0, 4, 8 and 12
+   * are the only possible values for nleft.
+   */
+  switch (nleft){
   case 0:
     break;
     
-  case 1:
+  case 4:
     acc0 = vec_madd(a0, b0, acc0);
     break;
 
-  case 2:
+  case 8:
     a1 = vec_perm(p1, p2, lvsl_a);
-    b1 = vec_ld(1*VS, b);
+    b1 = vec_ld(r1vs, b);
     acc0 = vec_madd(a0, b0, acc0);
     acc1 = vec_madd(a1, b1, acc1);
     break;
 
-  case 3:
+  case 12:
     a1 = vec_perm(p1, p2, lvsl_a);
-    b1 = vec_ld(1*VS, b);
+    b1 = vec_ld(r1vs, b);
     acc0 = vec_madd(a0, b0, acc0);
     a2 = vec_perm(p2, p3, lvsl_a);
-    b2 = vec_ld(2*VS, b);
+    b2 = vec_ld(r2vs, b);
     acc1 = vec_madd(a1, b1, acc1);
     acc2 = vec_madd(a2, b2, acc2);
     break;
-
-  default:
-    assert(0);
-    break;
   }
            
   acc0 = acc0 + acc1;
@@ -264,6 +157,7 @@
 }
 
 #endif
+}
 
 gr_fir_fff_vmx::gr_fir_fff_vmx()
   : gr_fir_fff_generic(),
@@ -290,7 +184,7 @@
 gr_fir_fff_vmx::set_taps(const std::vector<float> &inew_taps)
 {
   gr_fir_fff_generic::set_taps(inew_taps);     // call superclass
-  d_naligned_taps = gr_round_up(ntaps(), FLOATS_PER_VEC);
+  d_naligned_taps = gr_p2_round_up(ntaps(), FLOATS_PER_VEC);
 
   if (d_aligned_taps){
     free(d_aligned_taps);

Modified: 
gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/general/gr_math.h
===================================================================
--- gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/general/gr_math.h 
2008-07-22 00:21:02 UTC (rev 8967)
+++ gnuradio/branches/developers/eb/vmx/gnuradio-core/src/lib/general/gr_math.h 
2008-07-22 03:43:45 UTC (rev 8968)
@@ -174,4 +174,48 @@
   return gr_branchless_quad_45deg_slicer(x.real(), x.imag());
 }
 
+/*!
+ * \param x any value
+ * \param pow2 must be a power of 2
+ * \returns \p x rounded down to a multiple of \p pow2.
+ */
+static inline size_t
+gr_p2_round_down(size_t x, size_t pow2)
+{
+  return x & -pow2;
+}
+
+/*!
+ * \param x any value
+ * \param pow2 must be a power of 2
+ * \returns \p x rounded up to a multiple of \p pow2.
+ */
+static inline size_t
+gr_p2_round_up(size_t x, size_t pow2)
+{
+  return gr_p2_round_down(x + pow2 - 1, pow2);
+}
+
+/*!
+ * \param x any value
+ * \param pow2 must be a power of 2
+ * \returns \p x modulo \p pow2.
+ */
+static inline size_t
+gr_p2_modulo(size_t x, size_t pow2)
+{
+  return x & (pow2 - 1);
+}
+
+/*!
+ * \param x any value
+ * \param pow2 must be a power of 2
+ * \returns \p pow2 - (\p x modulo \p pow2).
+ */
+static inline size_t
+gr_p2_modulo_neg(size_t x, size_t pow2)
+{
+  return pow2 - gr_p2_modulo(x, pow2);
+}
+
 #endif /* _GR_MATH_H_ */