Re: [PATCH] PowerPC: modf optimization

[Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>]

On 03/26/2013 09:51 AM, Siddhesh Poyarekar wrote:
> On 25 March 2013 23:55, Adhemerval Zanella <azanella@linux.vnet.ibm.com> wrote:
>> I measuring with a simple benchmark that calls the function multiple times and
>> calculate the average time on different kind of inputs. I'll check the benchmark
>> testsuite.
>>
> Oh yes please!  The instructions in benchtests/Makefile should be
> sufficient, but please don't hesitate to reach out to me directly too
> if needed.
>
> Thanks,
> Siddhesh

Ok, here it an updated version using the benchtest framework. Using it I got:

PATCH - m32: modf: ITERS:200000: TOTAL:0.00169022s, MAX:16.982ns, MIN:7.118ns, 1.18328e+08 iter/s
MASTER:      modf: ITERS:200000: TOTAL:0.00307463s, MAX:18.642ns, MIN:13.8ns,  6.50484e+07 iter/s

PATCH - m64: modf: ITERS:200000: TOTAL:0.0011649s,  MAX:8.36ns, MIN:5.458ns,   1.71688e+08 iter/s
MASTER       modf: ITERS:200000: TOTAL:0.00340596s, MAX:21.344ns, MIN:16.85ns, 5.87206e+07 iter/s

Current patch adds the objects s_floor, s_ceil, s_floorf, and s_ceilf on libc.so
(since __ceil and __floor needs redirection to its implementation for POWER4). I
thought about an option on adding this optimization solely for POWER5+, avoiding
this inclusion. Any tips, advices, comments?

--

2013-03-25  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>

	* sysdeps/powerpc/fpu/s_modf.c: New file: modf optimization for POWER.
 	* sysdeps/powerpc/fpu/s_modff.c: New file: modff optimization for POWER.
	* sysdeps/powerpc/fpu/Makefile: Add s_floor[f] and s_ceil[f] for 
	powerpc64/POWER4 builds.
	* benchtests/Makefile: Add modf testcase.
	* benchtests/bench-modf.c: New file: Benchmark test for modf.

--

diff --git a/benchtests/Makefile b/benchtests/Makefile
index 74938b9..9c62d34 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -43,7 +43,7 @@
 #   See pow-inputs for an example.
 
 subdir := benchtests
-bench := exp pow
+bench := exp pow modf
 
 exp-ITER = 100000
 exp-ARGLIST = double
@@ -55,5 +55,10 @@ pow-ARGLIST = double:double
 pow-RET = double
 LDFLAGS-bench-pow = -lm
 
+modf-ITER = 100000
+modf-ARGLIST = double:"double*"
+modf-RET = double
+LDFLAGS-bench-modf = -lm
+
 include ../Makeconfig
 include ../Rules
diff --git a/benchtests/bench-modf.c b/benchtests/bench-modf.c
new file mode 100644
index 0000000..1751bee
--- /dev/null
+++ b/benchtests/bench-modf.c
@@ -0,0 +1,10 @@
+extern double modf ( double, double*);
+#define CALL_BENCH_FUNC(j, i) modf ( in[j].arg0, &i);
+struct args { double arg0; } in[] = {{42.42}, {-42.42}
+};
+#define NUM_SAMPLES (sizeof (in) / sizeof (struct args))
+static volatile double ret = 0.0;
+#define BENCH_FUNC(j) ({double iptr; ret =  CALL_BENCH_FUNC (j, iptr);})
+#define ITER 100000
+#define FUNCNAME "modf"
+#include "bench-skeleton.c"
diff --git a/sysdeps/powerpc/fpu/Makefile b/sysdeps/powerpc/fpu/Makefile
index fda59f9..fe1eadc 100644
--- a/sysdeps/powerpc/fpu/Makefile
+++ b/sysdeps/powerpc/fpu/Makefile
@@ -1,6 +1,8 @@
 ifeq ($(subdir),math)
 libm-support += fenv_const fe_nomask fe_mask t_sqrt
 
+# modf needs floof and ceil for powerpc64/POWER4
+sysdep_routines += s_floor s_ceil s_floorf s_ceilf
 # libm needs ld.so to access dl_hwcap
 $(objpfx)libm.so: $(elfobjdir)/ld.so
 endif
diff --git a/sysdeps/powerpc/fpu/s_modf.c b/sysdeps/powerpc/fpu/s_modf.c
new file mode 100644
index 0000000..b45bf66
--- /dev/null
+++ b/sysdeps/powerpc/fpu/s_modf.c
@@ -0,0 +1,58 @@
+/* Copyright (C) 2013 Free Software Foundation, Inc.
+   This file is part of the GNU C Library
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If
+   not, see <http://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <math_private.h>
+#include <math_ldbl_opt.h>
+
+double
+__modf (double x, double *iptr)
+{
+  if (__builtin_isinf (x))
+    {
+      *iptr = x;
+      return __copysign (0.0, x);
+    }
+  else if (__builtin_isnan (x))
+    {
+      *iptr = NAN;
+      return NAN;
+    }
+
+  if (x >= 0.0)
+    {
+      *iptr = __floor (x);
+      return (x - *iptr);
+    }
+  else
+    {
+      *iptr = __ceil (x);
+      return (x - *iptr);
+    }
+}
+weak_alias (__modf, modf)
+#ifdef NO_LONG_DOUBLE
+strong_alias (__modf, __modfl)
+weak_alias (__modf, modfl)
+#endif
+#ifdef IS_IN_libm
+# if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
+compat_symbol (libm, __modf, modfl, GLIBC_2_0);
+# endif
+#elif LONG_DOUBLE_COMPAT(libc, GLIBC_2_0)
+compat_symbol (libc, __modf, modfl, GLIBC_2_0);
+#endif
diff --git a/sysdeps/powerpc/fpu/s_modff.c b/sysdeps/powerpc/fpu/s_modff.c
new file mode 100644
index 0000000..55759cd
--- /dev/null
+++ b/sysdeps/powerpc/fpu/s_modff.c
@@ -0,0 +1,46 @@
+/* Copyright (C) 2013 Free Software Foundation, Inc.
+   This file is part of the GNU C Library
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If
+   not, see <http://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <math_private.h>
+
+float
+__modff (float x, float *iptr)
+{
+  if (__builtin_isinff (x))
+    {
+      *iptr = x;
+      return __copysignf (0.0, x);
+    }
+  else if (__builtin_isnanf (x))
+    {
+      *iptr = NAN;
+      return NAN;
+    }
+
+  if (x >= 0.0)
+    {
+      *iptr = __floorf (x);
+      return (x - *iptr);
+    }
+  else
+    {
+      *iptr = __ceilf (x);
+      return (x - *iptr);
+    }
+}
+weak_alias (__modff, modff)