This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] PowerPC - ilogb[f|l] optimization for POWER7
- From: Andreas Jaeger <aj at suse dot com>
- To: libc-alpha at sourceware dot org, Ryan Arnold <rsa at us dot ibm dot com>
- Cc: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- Date: Tue, 8 May 2012 15:12:45 +0200
- Subject: Re: [PATCH] PowerPC - ilogb[f|l] optimization for POWER7
- References: <4FA91281.8090109@linux.vnet.ibm.com>
On Tuesday, May 08, 2012 14:33:05 Adhemerval Zanella wrote:
> This patch provides optimized ilogb (60% on PPC32 and 20% PPC64),
> ilogbf (60% on PPC32 and 50% on PPC64), and ilogbl (3% on PPC32
> and 8% on PPC64). The optimization is done by avoiding float-point
> to integer transformation and by using VSX float-point bitwise
> instructions.
Ryan, please review this as well.
It looks ok to me, just noticed a few nits below.
> ---
>
> 2012-05-08 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
>
> * sysdeps/powerpc/powerpc32/power7/fpu/w_ilogb.c: New file: optimized
> ilogb for POWER7.
> * sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbf.c: New file: optimized
> ilogbf for POWER7.
> * sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbl.c: New file: optimized
> ilogbl for POWER7.
> * sysdeps/powerpc/powerpc64/power7/fpu/w_ilogb.c: New file: wrapper
> for the optimized logb for PPC64.
> * sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbf.c: New file: wrapper
> for the optimized logbf for PPC64.
> * sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbl.c: New file: wrapper
> for the optimized logbl for PPC64.
> * sysdeps/powerpc/powerpc32/power7/fpu/e_ilogb.c: New file: black file
> to avoid compilation of default implementation.
> * sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbf.c: Likewise.
> * sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbl.c: Likewise.
> * sysdeps/powerpc/powerpc64/power7/fpu/e_ilogb.c: Likewise.
> * sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbf.c: Likewise.
> * sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbl.c: Likewise.
> * sysdeps/powerpc/fpu/math_private.h: Add Optimized double to word
"optimized"
> cast for POWER7.
> * math/libm-test.inc (ilogb_test): New ilogb tests.
>
> diff --git a/math/libm-test.inc b/math/libm-test.inc
> index 542131d..677e5e9 100644
> --- a/math/libm-test.inc
> +++ b/math/libm-test.inc
> @@ -4082,6 +4082,10 @@ ilogb_test (void)
> TEST_f_i (ilogb, M_El, 1);
> TEST_f_i (ilogb, 1024, 10);
> TEST_f_i (ilogb, -2000, 10);
> + TEST_f_i (ilogb, 1.701412e+38, 127);
> +#ifndef TEST_FLOAT
> + TEST_f_i (ilogb, 8.988466e+307, 1023);
> +#endif
>
> /* ilogb (0.0) == FP_ILOGB0 plus invalid exception */
> errno = 0;
> diff --git a/sysdeps/powerpc/fpu/math_private.h
> b/sysdeps/powerpc/fpu/math_private.h index a916be3..aa3053a 100644
> --- a/sysdeps/powerpc/fpu/math_private.h
> +++ b/sysdeps/powerpc/fpu/math_private.h
> @@ -25,6 +25,29 @@
> #include <dl-procinfo.h>
> #include_next <math_private.h>
>
> +#if defined(_ARCH_PWR7)
> +
> +/* Optimized double to word cast for POWER7: the 'ori 2,2,0'
> + instructions between the store double / load integer is
> + to force a new dispatch group. */
> +#undef DOUBLE_TO_WORDS
> +#define DOUBLE_TO_WORDS(d, i) \
> + do { \
> + double d__ = d; \
> + int32_t i__; \
> + ieee_double_shape_type iw_u; \
> + __asm ( \
> + "fctiwz %1,%1\n" \
> + "stfd %1,%2\n" \
> + "ori 2,2,0\n" \
> + "lwz %0,%3\n" \
> + : "=r" (i__) \
> + : "f" (d__), "m" (iw_u.value), "m" (iw_u.word)); \
> + i = i__; \
> + } while (0)
> +
> +#endif /* __ARCH_PWR7 */
> +
> # if __WORDSIZE == 64 || defined _ARCH_PWR4
> # define __CPU_HAS_FSQRT 1
> # else
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogb.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogb.c new file mode 100644
> index 0000000..34f1a9b
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogb.c
> @@ -0,0 +1 @@
> +/* ilogb implementation is at w_ilogb.c */
I think it's "in" instead of "at" - but I'm not a native speaker either.
Change it below as well.
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbf.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbf.c new file mode 100644
> index 0000000..1bbe157
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbf.c
> @@ -0,0 +1 @@
> +/* ilogbf implementation is at w_ilogbf.c */
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbl.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbl.c new file mode 100644
> index 0000000..7684390
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/e_ilogbl.c
> @@ -0,0 +1 @@
> +/* ilogbl implementation is at w_ilogbl.c */
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogb.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogb.c new file mode 100644
> index 0000000..525f76b
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogb.c
> @@ -0,0 +1,83 @@
> +/* ilogb(). PowerPC64/POWER7 version.
> + Copyright (C) 2012 Free Software Foundation, Inc.
> + Contributed by Adhemerval Zanella Netto <azanella@br.ibm.com>.
> + 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 Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 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
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <http://www.gnu.org/licenses/>. */
> +
> +#include <fenv.h>
> +#include <errno.h>
> +#include <math.h>
> +#include <math_private.h>
> +
> +
> +/* This implementation avoid FP to INT conversions by using VSX bitwise
> + * instructions over FP values. */
"avoids"
> +
> +static const double two1div52 = 2.220446049250313e-16; /* 1/2**52 */
> +static const double two10m1 = -1023.0; /* 2**10 -1 */
> +
> +static const union {
> + unsigned long long mask;
> + double d;
> +}
> +exp_mask = { 0x7ff0000000000000ULL },
> +sig_mask = { 0x000FFFFFFFFFFFFFULL };
> +
> +int
> +__ilogb (double x)
> +{
> + double dret;
> + int ret;
> +
> + if (__builtin_expect (x == 0.0, 0))
> + {
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + return FP_ILOGB0;
> + }
> + /* dret = x & 0x7ff0000000000000; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + "fcfid %0,%0"
> + : "=f" (dret)
> + : "f" (x), "f" (exp_mask.d));
> + /* dret = (x >> 52) - 1023.0; */
> + dret = (dret * two1div52) + two10m1;
> + if (__builtin_expect (dret > -two10m1, 0))
> + {
> + double sig;
> + /* sig = x & 0x000FFFFFFFFFFFFF; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + : "=f" (sig)
> + : "f" (x), "f" (sig_mask.d));
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + /* NaN differs by INF by having some bits set in significand. */
> + if (sig > 0.0)
> + return FP_ILOGBNAN;
> + return INT_MAX;
> + }
> + DOUBLE_TO_WORDS (dret, ret);
> + return ret;
> +}
> +
> +weak_alias (__ilogb, ilogb)
> +
> +#ifdef NO_LONG_DOUBLE
> +strong_alias (__ilogb, __ilogbl)
> +weak_alias (__ilogb, ilogbl)
> +#endif
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbf.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbf.c new file mode 100644
> index 0000000..a8b56e5
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbf.c
> @@ -0,0 +1,78 @@
> +/* ilogbf(). PowerPC64/POWER7 version.
> + Copyright (C) 2012 Free Software Foundation, Inc.
> + Contributed by Adhemerval Zanella Netto <azanella@br.ibm.com>.
> + 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 Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 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
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <http://www.gnu.org/licenses/>. */
> +
> +#include <fenv.h>
> +#include <errno.h>
> +#include <math.h>
> +#include <math_private.h>
> +
> +
> +/* This implementation avoid FP to INT conversions by using VSX bitwise
> + * instructions over FP values. */
"avoids"
> +
> +static const double two1div52 = 2.220446049250313e-16; /* 1/2**52 */
> +static const double two10m1 = -1023.0; /* -2**10 + 1 */
> +static const double two7m1 = -127.0; /* -2**7 + 1 */
> +
> +static const union {
> + unsigned long long mask;
> + double d;
> +}
> +exp_mask = { 0x7ff0000000000000ULL },
> +sig_mask = { 0x000FFFFFFFFFFFFFULL };
> +
> +int
> +__ilogbf (float x)
> +{
> + double dret;
> + int ret;
> +
> + if (__builtin_expect (x == 0.0, 0))
> + {
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + return FP_ILOGB0;
> + }
> + /* dret = x & 0x7ff0000000000000; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + "fcfid %0,%0"
> + : "=f" (dret)
> + : "f" (x), "f" (exp_mask.d));
> + /* dret = (x >> 52) - 1023.0; */
> + dret = (dret * two1div52) + two10m1;
> + if (__builtin_expect (dret > -two7m1, 0))
> + {
> + double sig;
> + /* sig = x & 0x000FFFFFFFFFFFFF; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + : "=f" (sig)
> + : "f" (x), "f" (sig_mask.d));
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + /* NaN differs by INF by having some bits set in significand. */
> + if (sig > 0.0)
> + return FP_ILOGBNAN;
> + return INT_MAX;
> + }
> + DOUBLE_TO_WORDS (dret, ret);
> + return ret;
> +}
> +weak_alias (__ilogbf, ilogbf)
> diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbl.c
> b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbl.c new file mode 100644
> index 0000000..7563020
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbl.c
> @@ -0,0 +1,81 @@
> +/* ilogb(). PowerPC64/POWER7 version.
> + Copyright (C) 2012 Free Software Foundation, Inc.
> + Contributed by Adhemerval Zanella Netto <azanella@br.ibm.com>.
> + 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 Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 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
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <http://www.gnu.org/licenses/>. */
> +
> +#include <fenv.h>
> +#include <errno.h>
> +#include <math.h>
> +#include <math_private.h>
> +#include <math_ldbl_opt.h>
> +
> +/* This implementation avoid FP to INT conversions by using VSX bitwise
> + * instructions over FP values. */
"avoids"
> +
> +static const double two1div52 = 2.220446049250313e-16; /* 1/2**52 */
> +static const double two10m1 = -1023.0; /* 2**10 -1 */
> +
> +static const union {
> + unsigned long long mask;
> + double d;
> +}
> +exp_mask = { 0x7ff0000000000000ULL },
> +sig_mask = { 0x000FFFFFFFFFFFFFULL };
> +
> +int
> +__ilogbl (long double x)
> +{
> + double xh, xl;
> + double dret;
> + int ret;
> +
> + if (__builtin_expect (x == 0.0L, 0))
> + {
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + return FP_ILOGB0;
> + }
> +
> + ldbl_unpack (x, &xh, &xl);
> + /* dret = x & 0x7ff0000000000000; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + "fcfid %0,%0"
> + : "=f" (dret)
> + : "f" (xh), "f" (exp_mask.d));
> + /* dret = (x >> 52) - 1023.0; */
> + dret = (dret * two1div52) + two10m1;
> + if (__builtin_expect (dret > -two10m1, 0))
> + {
> + double sig;
> + /* sig = x & 0x000FFFFFFFFFFFFF; */
> + asm (
> + "xxland %x0,%x1,%x2\n"
> + : "=f" (sig)
> + : "f" (x), "f" (sig_mask.d));
> + feraiseexcept (FE_INVALID);
> + __set_errno (EDOM);
> + /* NaN differs by INF by having some bits set in significand. */
> + if (sig > 0.0)
> + return FP_ILOGBNAN;
> + return INT_MAX;
> + }
> + DOUBLE_TO_WORDS (dret, ret);
> + return ret;
> +}
> +
> +long_double_symbol (libm, __ilogbl, ilogbl);
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogb.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogb.c new file mode 100644
> index 0000000..34f1a9b
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogb.c
> @@ -0,0 +1 @@
> +/* ilogb implementation is at w_ilogb.c */
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbf.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbf.c new file mode 100644
> index 0000000..1bbe157
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbf.c
> @@ -0,0 +1 @@
> +/* ilogbf implementation is at w_ilogbf.c */
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbl.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbl.c new file mode 100644
> index 0000000..7684390
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/e_ilogbl.c
> @@ -0,0 +1 @@
> +/* ilogbl implementation is at w_ilogbl.c */
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogb.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogb.c new file mode 100644
> index 0000000..48a85ca
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogb.c
> @@ -0,0 +1 @@
> +#include <sysdeps/powerpc/powerpc32/power7/fpu/w_ilogb.c>
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbf.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbf.c new file mode 100644
> index 0000000..9220977
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbf.c
> @@ -0,0 +1 @@
> +#include <sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbf.c>
> diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbl.c
> b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbl.c new file mode 100644
> index 0000000..bd76c6e
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/power7/fpu/w_ilogbl.c
> @@ -0,0 +1 @@
> +#include <sysdeps/powerpc/powerpc32/power7/fpu/w_ilogbl.c>
--
Andreas Jaeger aj@{suse.com,opensuse.org} Twitter/Identica: jaegerandi
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn,Jennifer Guild,Felix Imendörffer,HRB16746 (AG Nürnberg)
GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126