Re: [PATCH v2] powerpc: Optimized conversion of IBM long double to int128/int64
On 10/25/2018 09:40 PM, Joseph Myers wrote:
On Thu, 25 Oct 2018, Rajalakshmi Srinivasaraghavan wrote:
+ if (__builtin_unpack_longdouble (a, 0) < TWO53)
+{
+ /* In this case the integer portion is completely contained
+ within the high double. So use the hardware convert to
+ integer doubleword, and then extend to int. */
+ l1 = __builtin_unpack_longdouble (a, 0);
+ result = l1;
But if the high double is a positive integer, and the low double is
strictly negative (not -0), you need to subtract 1 to get a result that's
correctly truncated towards zero.
Do you suggest like this?
if (__builtin_unpack_longdouble (a, 0) < TWO53)
{
l1 = __builtin_unpack_longdouble (a, 0);
if (__builtin_unpack_longdouble (a, 1) < 0.0)
result = l1 - 1;
else
result = l1;
}
+ if (__builtin_unpack_longdouble (a, 0) < TWO53)
+{
+ /* In this case the integer portion is completely contained
+ within the high double. So use the hardware convert to
+ integer doubleword, and then extend to __int128. */
+ l1 = __builtin_unpack_longdouble (a, 0);
+ result = l1;
Likewise here.
--
Thanks
Rajalakshmi S
Re: [PATCH] powerpc: Optimized conversion of IBM long double to int128/int64
On 10/23/2018 10:07 PM, Segher Boessenkool wrote: On Tue, Oct 23, 2018 at 09:01:26PM +0530, Rajalakshmi Srinivasaraghavan wrote: This new implementation of fixunstfdi and fixunstfti gives 16X performance improvement. :-) * libgcc/config/rs6000/t-ppc64-fp (LIB2ADD): Add $(srcdir)/config/rs6000/fixunstfti.c. And fixunstfdi.c? Added now. * libgcc/config/rs6000/ppc64-fp.c (__fixunstfdi): Remove definition. * libgcc/config/rs6000/fixunstfti.c: New file. * libgcc/config/rs6000/fixunstfdi.c: Likewise. * libgcc/config/rs6000/ibm-ldouble.h: Likewise. libgcc/ has its own changelog; the path names in the changelog should be relative to that (so should start with config/). Fixed in the v2 patch. Does -m32 still work after this? (Did it before?) __int128 is not supported on powerpc m32 build. I tried to build gcc on powerpc 32bit machine to make sure there are no build failures. Okay for trunk with the changelog fixed up (unless it broke -m32 ;-) ) Segher -- Thanks Rajalakshmi S
[PATCH v2] powerpc: Optimized conversion of IBM long double to int128/int64
This new implementation of fixunstfdi and fixunstfti
gives 16X performance improvement.
The design is focused on:
- Making sure the end result was a pure leaf function that
only needed builtins or inline functions.
- Assumed power8 direct register transfer and accessed the IBM
long double as int bit field structure.
- Understanding the quirks of IBM long double and decompose the
code in to a set of optimized sub cases.
Tested on powerpc64le.
libgcc ChangeLog:
2018-10-25 Steven Munroe
Rajalakshmi Srinivasaraghavan
* config/rs6000/t-ppc64-fp (LIB2ADD): Add
$(srcdir)/config/rs6000/fixunstfti.c and
$(srcdir)/config/rs6000/fixunstfdi.c.
* config/rs6000/ppc64-fp.c (__fixunstfdi): Remove definition.
* config/rs6000/fixunstfti.c: New file.
* config/rs6000/fixunstfdi.c: Likewise.
* config/rs6000/ibm-ldouble.h: Likewise.
---
libgcc/config/rs6000/fixunstfdi.c | 124
libgcc/config/rs6000/fixunstfti.c | 125 +
libgcc/config/rs6000/ibm-ldouble.h | 121
libgcc/config/rs6000/ppc64-fp.c| 24 --
libgcc/config/rs6000/t-ppc64-fp| 5 +-
5 files changed, 374 insertions(+), 25 deletions(-)
create mode 100755 libgcc/config/rs6000/fixunstfdi.c
create mode 100755 libgcc/config/rs6000/fixunstfti.c
create mode 100755 libgcc/config/rs6000/ibm-ldouble.h
diff --git a/libgcc/config/rs6000/fixunstfdi.c
b/libgcc/config/rs6000/fixunstfdi.c
new file mode 100755
index 000..1b1a4f280bd
--- /dev/null
+++ b/libgcc/config/rs6000/fixunstfdi.c
@@ -0,0 +1,124 @@
+/* Convert IBM long double to 64bit unsigned integer.
+
+ Copyright (C) 2018 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.
+
+ In addition to the permissions in the GNU Lesser General Public
+ License, the Free Software Foundation gives you unlimited
+ permission to link the compiled version of this file into
+ combinations with other programs, and to distribute those
+ combinations without any restriction coming from the use of this
+ file. (The Lesser General Public License restrictions do apply in
+ other respects; for example, they cover modification of the file,
+ and distribution when not linked into a combine executable.)
+
+ 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/>. */
+
+#if defined(__powerpc64__) || defined (__64BIT__) || defined(__ppc64__)
+#include
+#include "ibm-ldouble.h"
+
+typedef unsigned int UDItype __attribute__ ((mode (DI)));
+typedef float TFtype __attribute__ ((mode (TF)));
+extern UDItype __fixunstfdi (TFtype);
+
+#define TWO53 9007199254740992.0L
+#define TWO64 18446744073709551616.0L
+
+UDItype
+__fixunstfdi (TFtype a)
+{
+ unsigned long result;
+ unsigned long qi0, qi1;
+ union ibm_extended_long_double ld;
+ uint64_t l0, l1;
+ long exp0, exp1;
+ const uint64_t two52 = 0x10;
+ if (__builtin_unpack_longdouble (a, 0) < TWO53)
+{
+ /* In this case the integer portion is completely contained
+ within the high double. So use the hardware convert to
+ integer doubleword, and then extend to int. */
+ l1 = __builtin_unpack_longdouble (a, 0);
+ result = l1;
+}
+ else
+{
+ if (a < TWO64)
+ {
+ ld.ld = a;
+ l0 = two52 | ((uint64_t)ld.d[0].ieee.mantissa0 << 32)
+ | ld.d[0].ieee.mantissa1;
+ l1 = two52 | ((uint64_t)ld.d[1].ieee.mantissa0 << 32)
+ | ld.d[1].ieee.mantissa1;
+ exp0 = ld.d[0].ieee.exponent - IEEE754_DOUBLE_BIAS;
+ exp1 = ld.d[1].ieee.exponent - IEEE754_DOUBLE_BIAS;
+ /* The high double shift is (non-negative) because in this
+case we know the value it greater than 2^53 -1. */
+ qi0 = l0;
+ qi0 = qi0 << (exp0 - 52);
+ /* The low double is tricky because it could be
+zero/denormal and have a large negative exponent. */
+ if ( exp1 > -1022)
+ {
+ /* Need to right justify the integer portion of the
+low double. This may be a left or right shift. */
+ exp1 = exp1 - 52;
+ if (exp1 < 0)
+ {
+ /* Negative exponent, shift right to tru
[PATCH] powerpc: Optimized conversion of IBM long double to int128/int64
This new implementation of fixunstfdi and fixunstfti
gives 16X performance improvement.
The design is focused on:
- Making sure the end result was a pure leaf function that
only needed builtins or inline functions.
- Assumed power8 direct register transfer and accessed the IBM
long double as int bit field structure.
- Understanding the quirks of IBM long double and decompose the
code in to a set of optimized sub cases.
Tested on powerpc64le.
2018-10-20 Steven Munroe
Rajalakshmi Srinivasaraghavan
* libgcc/config/rs6000/t-ppc64-fp (LIB2ADD): Add
$(srcdir)/config/rs6000/fixunstfti.c.
* libgcc/config/rs6000/ppc64-fp.c (__fixunstfdi): Remove definition.
* libgcc/config/rs6000/fixunstfti.c: New file.
* libgcc/config/rs6000/fixunstfdi.c: Likewise.
* libgcc/config/rs6000/ibm-ldouble.h: Likewise.
---
libgcc/config/rs6000/fixunstfdi.c | 124
libgcc/config/rs6000/fixunstfti.c | 125 +
libgcc/config/rs6000/ibm-ldouble.h | 121
libgcc/config/rs6000/ppc64-fp.c| 24 --
libgcc/config/rs6000/t-ppc64-fp| 5 +-
5 files changed, 374 insertions(+), 25 deletions(-)
create mode 100755 libgcc/config/rs6000/fixunstfdi.c
create mode 100755 libgcc/config/rs6000/fixunstfti.c
create mode 100755 libgcc/config/rs6000/ibm-ldouble.h
diff --git a/libgcc/config/rs6000/fixunstfdi.c
b/libgcc/config/rs6000/fixunstfdi.c
new file mode 100755
index 000..1b1a4f280bd
--- /dev/null
+++ b/libgcc/config/rs6000/fixunstfdi.c
@@ -0,0 +1,124 @@
+/* Convert IBM long double to 64bit unsigned integer.
+
+ Copyright (C) 2018 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.
+
+ In addition to the permissions in the GNU Lesser General Public
+ License, the Free Software Foundation gives you unlimited
+ permission to link the compiled version of this file into
+ combinations with other programs, and to distribute those
+ combinations without any restriction coming from the use of this
+ file. (The Lesser General Public License restrictions do apply in
+ other respects; for example, they cover modification of the file,
+ and distribution when not linked into a combine executable.)
+
+ 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/>. */
+
+#if defined(__powerpc64__) || defined (__64BIT__) || defined(__ppc64__)
+#include
+#include "ibm-ldouble.h"
+
+typedef unsigned int UDItype __attribute__ ((mode (DI)));
+typedef float TFtype __attribute__ ((mode (TF)));
+extern UDItype __fixunstfdi (TFtype);
+
+#define TWO53 9007199254740992.0L
+#define TWO64 18446744073709551616.0L
+
+UDItype
+__fixunstfdi (TFtype a)
+{
+ unsigned long result;
+ unsigned long qi0, qi1;
+ union ibm_extended_long_double ld;
+ uint64_t l0, l1;
+ long exp0, exp1;
+ const uint64_t two52 = 0x10;
+ if (__builtin_unpack_longdouble (a, 0) < TWO53)
+{
+ /* In this case the integer portion is completely contained
+ within the high double. So use the hardware convert to
+ integer doubleword, and then extend to int. */
+ l1 = __builtin_unpack_longdouble (a, 0);
+ result = l1;
+}
+ else
+{
+ if (a < TWO64)
+ {
+ ld.ld = a;
+ l0 = two52 | ((uint64_t)ld.d[0].ieee.mantissa0 << 32)
+ | ld.d[0].ieee.mantissa1;
+ l1 = two52 | ((uint64_t)ld.d[1].ieee.mantissa0 << 32)
+ | ld.d[1].ieee.mantissa1;
+ exp0 = ld.d[0].ieee.exponent - IEEE754_DOUBLE_BIAS;
+ exp1 = ld.d[1].ieee.exponent - IEEE754_DOUBLE_BIAS;
+ /* The high double shift is (non-negative) because in this
+case we know the value it greater than 2^53 -1. */
+ qi0 = l0;
+ qi0 = qi0 << (exp0 - 52);
+ /* The low double is tricky because it could be
+zero/denormal and have a large negative exponent. */
+ if ( exp1 > -1022)
+ {
+ /* Need to right justify the integer portion of the
+low double. This may be a left or right shift. */
+ exp1 = exp1 - 52;
+ if (exp1 < 0)
+ {
+ /* Negative exponent, shift right to truncate. */
+
