This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] PowerPC - logb[f|l] optimization for POWER7
On 05/08/2012 12:45 PM, Ryan S. Arnold wrote:
> No colon necessary after "New file:". It should be "New file.
> Optimize logb for POWER7".
> I prefer that this mention that it #includes the powerpc32/power7/ .c
> file. Something like:
>
> "New file. Use powerpc32/power7/logb[fl].c via #include.
ChangeLog following this updates now.
> Roland has indicated that he no longer wants "Contributed by"
> statements.. The git log will serve as attribution. This applies to
> all files in this patchset.
I already removed the "Contributed by" lines.
> I would like if there were a comment indicating what's going on, i.e.,
> /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF]. */
>
>
> Once again, please comment the special case with an indication of
> what's being returned (if it is a special value indicated by the
> spec). Please indicate the special value in the comment rather than
> just a magic number that represents it. This applies to all files in
> this patchset.
I added comments for why the return operations are used for.
>> + return ret == -0.0 ? 0.0 : ret;
> Is this faster than an builtin abs call?
For POWER7, the compiler will optimize this and by replacing a '0.0' for a '__builtin_abs (ret)' will
just trade a 'xxlor' instruction by 'xsabsdp'. It is not a faster, both have similar latency.
> Should be "Since operations are done with double we don't need additional..."
Fixed.
---
2012-05-09 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
* sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c: New file. Optimized
logb for POWER7.
* sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c: New file. Optimized
logbf for POWER7.
* sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c: New file. Optimized
logbl for POWER7.
* sysdeps/powerpc/powerpc64/power7/fpu/s_logb.c: New file. Use
powerpc32/power7/fpu/s_logb.c via #include.
* sysdeps/powerpc/powerpc64/power7/fpu/s_logbf.c: New file. Use
powerpc32/power7/fpu/s_logbf.c via #include.
* sysdeps/powerpc/powerpc64/power7/fpu/s_logbl.c: New file. Use
powerpc32/power7/fpu/s_logbl.c via #include.
diff --git a/math/libm-test.inc b/math/libm-test.inc
index 542131d..5a38dbf 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -5376,6 +5376,22 @@ logb_test (void)
TEST_f_f (logb, 1024, 10);
TEST_f_f (logb, -2000, 10);
+ TEST_f_f (logb, 0x0.1p-127, -131);
+ TEST_f_f (logb, 0x0.01p-127, -135);
+ TEST_f_f (logb, 0x0.011p-127, -135);
+#ifndef TEST_FLOAT
+ TEST_f_f (logb, 0x0.8p-1022, -1023);
+ TEST_f_f (logb, 0x0.1p-1022, -1026);
+ TEST_f_f (logb, 0x0.00111p-1022, -1034);
+ TEST_f_f (logb, 0x0.00001p-1022, -1042);
+ TEST_f_f (logb, 0x0.000011p-1022, -1042);
+ TEST_f_f (logb, 0x0.0000000000001p-1022, -1074);
+#endif
+#if defined TEST_LDOUBLE && LDBL_MIN_EXP - LDBL_MANT_DIG <= -16400
+ TEST_f_f (logb, 0x1p-16400L, -16400);
+ TEST_f_f (logb, 0x.00000000001p-16382L, -16426);
+#endif
+
END (logb);
}
diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c
new file mode 100644
index 0000000..a5d21f4
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c
@@ -0,0 +1,75 @@
+/* logb(). PowerPC/POWER7 version.
+ Copyright (C) 2012 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 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 "math_private.h"
+
+/* This implementation avoid FP to INT conversions by using VSX bitwise
+ instructions over FP values. */
+
+static const double two1div52 = 2.220446049250313e-16; /* 1/2**52 */
+static const double two10m1 = -1023.0; /* 2**10 -1 */
+
+/* FP mask to extract the exponent. */
+static const union {
+ unsigned long long mask;
+ double d;
+} mask = { 0x7ff0000000000000ULL };
+
+double
+__logb (double x)
+{
+ double ret;
+
+ if (__builtin_expect (x == 0.0, 0))
+ /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF]. */
+ return -1.0 / __builtin_fabs (x);
+
+ /* ret = x & 0x7ff0000000000000; */
+ asm (
+ "xxland %x0,%x1,%x2\n"
+ "fcfid %0,%0"
+ : "=f" (ret)
+ : "f" (x), "f" (mask.d));
+ /* ret = (ret >> 52) - 1023.0; */
+ ret = (ret * two1div52) + two10m1;
+ if (__builtin_expect (ret > -two10m1, 0))
+ /* Multiplication is used to set logb (+-INF) = INF. */
+ return (x * x);
+ else if (__builtin_expect (ret == two10m1, 0))
+ {
+ /* POSIX specifies that denormal numbers are treated as
+ though they were normalized. */
+ int32_t lx, ix;
+ int m1, m2, ma;
+
+ EXTRACT_WORDS (ix , lx, x);
+ m1 = (ix == 0) ? 0 : __builtin_clz (ix);
+ m2 = (lx == 0) ? 0 : __builtin_clz (lx);
+ ma = (m1 == 0) ? m2 + 32 : m1;
+ return -1022.0 + (double)(11 - ma);
+ }
+ /* Test to avoid logb_downward (0.0) == -0.0. */
+ return ret == -0.0 ? 0.0 : ret;
+}
+
+weak_alias (__logb, logb)
+
+#ifdef NO_LONG_DOUBLE
+strong_alias (__logb, __logbl)
+weak_alias (__logb, logbl)
+#endif
diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c
new file mode 100644
index 0000000..33254b5
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c
@@ -0,0 +1,60 @@
+/* logbf(). PowerPC/POWER7 version.
+ Copyright (C) 2012 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 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 "math_private.h"
+
+/* This implementation avoid FP to INT conversions by using VSX bitwise
+ instructions over FP values. */
+
+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 */
+
+/* FP mask to extract the exponent. */
+static const union {
+ unsigned long long mask;
+ double d;
+} mask = { 0x7ff0000000000000ULL };
+
+float
+__logbf (float x)
+{
+ /* VSX operation are all done internally as double. */
+ double ret;
+
+ if (__builtin_expect (x == 0.0, 0))
+ /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF]. */
+ return -1.0 / __builtin_fabsf (x);
+
+ /* ret = x & 0x7f800000; */
+ asm (
+ "xxland %x0,%x1,%x2\n"
+ "fcfid %0,%0"
+ : "=f"(ret)
+ : "f" (x), "f" (mask.d));
+ /* ret = (ret >> 52) - 1023.0, since ret is double. */
+ ret = (ret * two1div52) + two10m1;
+ if (__builtin_expect (ret > -two7m1, 0))
+ /* Multiplication is used to set logb (+-INF) = INF. */
+ return (x * x);
+ /* Since operations are done with double we don't need
+ additional tests for subnormal numbers.
+ The test is to avoid logb_downward (0.0) == -0.0. */
+ return ret == -0.0 ? 0.0 : ret;
+}
+weak_alias (__logbf, logbf)
diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c
new file mode 100644
index 0000000..5dd0cda
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c
@@ -0,0 +1,72 @@
+/* logbl(). PowerPC/POWER7 version.
+ Copyright (C) 2012 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 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 <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. */
+
+static const double two1div52 = 2.220446049250313e-16; /* 1/2**52 */
+static const double two10m1 = -1023.0; /* 2**10 -1 */
+
+/* FP mask to extract the exponent. */
+static const union {
+ unsigned long long mask;
+ double d;
+} mask = { 0x7ff0000000000000ULL };
+
+long double
+__logbl (long double x)
+{
+ double xh, xl;
+ double ret;
+
+ if (__builtin_expect (x == 0.0L, 0))
+ /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF]. */
+ return -1.0L / __builtin_fabsl (x);
+
+ ldbl_unpack (x, &xh, &xl);
+ /* ret = x & 0x7ff0000000000000; */
+ asm (
+ "xxland %x0,%x1,%x2\n"
+ "fcfid %0,%0"
+ : "=f" (ret)
+ : "f" (xh), "f" (mask.d));
+ /* ret = (ret >> 52) - 1023.0; */
+ ret = (ret * two1div52) + two10m1;
+ if (__builtin_expect (ret > -two10m1, 0))
+ /* Multiplication is used to set logb (+-INF) = INF. */
+ return (xh * xh);
+ else if (__builtin_expect (ret == two10m1, 0))
+ {
+ int64_t lx, hx;
+ int m1, m2, ma;
+
+ GET_LDOUBLE_WORDS64 (hx, lx, x);
+ m1 = (hx == 0) ? 0 : __builtin_clzll (hx);
+ m2 = (lx == 0) ? 0 : __builtin_clzll (lx);
+ ma = (m1 == 0) ? m2 + 64 : m1;
+ return -1022.0 + (double)(11 - ma);
+ }
+ /* Test to avoid logb_downward (0.0) == -0.0. */
+ return ret == -0.0 ? 0.0 : ret;
+}
+
+long_double_symbol (libm, __logbl, logbl);
diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/s_logb.c b/sysdeps/powerpc/powerpc64/power7/fpu/s_logb.c
new file mode 100644
index 0000000..ff3a9e0
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power7/fpu/s_logb.c
@@ -0,0 +1 @@
+#include <sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c>
diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/s_logbf.c b/sysdeps/powerpc/powerpc64/power7/fpu/s_logbf.c
new file mode 100644
index 0000000..e79a28f
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power7/fpu/s_logbf.c
@@ -0,0 +1 @@
+#include <sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c>
diff --git a/sysdeps/powerpc/powerpc64/power7/fpu/s_logbl.c b/sysdeps/powerpc/powerpc64/power7/fpu/s_logbl.c
new file mode 100644
index 0000000..463e411
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power7/fpu/s_logbl.c
@@ -0,0 +1 @@
+#include <sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c>
--
1.6.0.2
--
Adhemerval Zanella Netto
Software Engineer
Linux Technology Center Brazil
Toolchain / GLIBC on Power Architecture
azanella@linux.vnet.ibm.com / azanella@br.ibm.com
+55 61 8642-9890