qemu/fpu/softfloat.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

5214 lines
147 KiB
C
Raw Normal View History

/*
* QEMU float support
*
2015-01-12 17:38:28 +03:00
* The code in this source file is derived from release 2a of the SoftFloat
* IEC/IEEE Floating-point Arithmetic Package. Those parts of the code (and
* some later contributions) are provided under that license, as detailed below.
* It has subsequently been modified by contributors to the QEMU Project,
* so some portions are provided under:
* the SoftFloat-2a license
* the BSD license
* GPL-v2-or-later
*
* Any future contributions to this file after December 1st 2014 will be
* taken to be licensed under the Softfloat-2a license unless specifically
* indicated otherwise.
*/
/*
===============================================================================
This C source file is part of the SoftFloat IEC/IEEE Floating-point
Arithmetic Package, Release 2a.
Written by John R. Hauser. This work was made possible in part by the
International Computer Science Institute, located at Suite 600, 1947 Center
Street, Berkeley, California 94704. Funding was partially provided by the
National Science Foundation under grant MIP-9311980. The original version
of this code was written as part of a project to build a fixed-point vector
processor in collaboration with the University of California at Berkeley,
overseen by Profs. Nelson Morgan and John Wawrzynek. More information
is available through the Web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
arithmetic/SoftFloat.html'.
THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort
has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO
PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
Derivative works are acceptable, even for commercial purposes, so long as
(1) they include prominent notice that the work is derivative, and (2) they
include prominent notice akin to these four paragraphs for those parts of
this code that are retained.
===============================================================================
*/
2015-01-12 17:38:28 +03:00
/* BSD licensing:
* Copyright (c) 2006, Fabrice Bellard
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its contributors
* may be used to endorse or promote products derived from this software without
* specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
* THE POSSIBILITY OF SUCH DAMAGE.
*/
/* Portions of this work are licensed under the terms of the GNU GPL,
* version 2 or later. See the COPYING file in the top-level directory.
*/
/* softfloat (and in particular the code in softfloat-specialize.h) is
* target-dependent and needs the TARGET_* macros.
*/
#include "qemu/osdep.h"
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
#include <math.h>
#include "qemu/bitops.h"
#include "fpu/softfloat.h"
/* We only need stdlib for abort() */
/*----------------------------------------------------------------------------
| Primitive arithmetic functions, including multi-word arithmetic, and
| division and square root approximations. (Can be specialized to target if
| desired.)
*----------------------------------------------------------------------------*/
#include "fpu/softfloat-macros.h"
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
/*
* Hardfloat
*
* Fast emulation of guest FP instructions is challenging for two reasons.
* First, FP instruction semantics are similar but not identical, particularly
* when handling NaNs. Second, emulating at reasonable speed the guest FP
* exception flags is not trivial: reading the host's flags register with a
* feclearexcept & fetestexcept pair is slow [slightly slower than soft-fp],
* and trapping on every FP exception is not fast nor pleasant to work with.
*
* We address these challenges by leveraging the host FPU for a subset of the
* operations. To do this we expand on the idea presented in this paper:
*
* Guo, Yu-Chuan, et al. "Translating the ARM Neon and VFP instructions in a
* binary translator." Software: Practice and Experience 46.12 (2016):1591-1615.
*
* The idea is thus to leverage the host FPU to (1) compute FP operations
* and (2) identify whether FP exceptions occurred while avoiding
* expensive exception flag register accesses.
*
* An important optimization shown in the paper is that given that exception
* flags are rarely cleared by the guest, we can avoid recomputing some flags.
* This is particularly useful for the inexact flag, which is very frequently
* raised in floating-point workloads.
*
* We optimize the code further by deferring to soft-fp whenever FP exception
* detection might get hairy. Two examples: (1) when at least one operand is
* denormal/inf/NaN; (2) when operands are not guaranteed to lead to a 0 result
* and the result is < the minimum normal.
*/
#define GEN_INPUT_FLUSH__NOCHECK(name, soft_t) \
static inline void name(soft_t *a, float_status *s) \
{ \
if (unlikely(soft_t ## _is_denormal(*a))) { \
*a = soft_t ## _set_sign(soft_t ## _zero, \
soft_t ## _is_neg(*a)); \
float_raise(float_flag_input_denormal, s); \
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
} \
}
GEN_INPUT_FLUSH__NOCHECK(float32_input_flush__nocheck, float32)
GEN_INPUT_FLUSH__NOCHECK(float64_input_flush__nocheck, float64)
#undef GEN_INPUT_FLUSH__NOCHECK
#define GEN_INPUT_FLUSH1(name, soft_t) \
static inline void name(soft_t *a, float_status *s) \
{ \
if (likely(!s->flush_inputs_to_zero)) { \
return; \
} \
soft_t ## _input_flush__nocheck(a, s); \
}
GEN_INPUT_FLUSH1(float32_input_flush1, float32)
GEN_INPUT_FLUSH1(float64_input_flush1, float64)
#undef GEN_INPUT_FLUSH1
#define GEN_INPUT_FLUSH2(name, soft_t) \
static inline void name(soft_t *a, soft_t *b, float_status *s) \
{ \
if (likely(!s->flush_inputs_to_zero)) { \
return; \
} \
soft_t ## _input_flush__nocheck(a, s); \
soft_t ## _input_flush__nocheck(b, s); \
}
GEN_INPUT_FLUSH2(float32_input_flush2, float32)
GEN_INPUT_FLUSH2(float64_input_flush2, float64)
#undef GEN_INPUT_FLUSH2
#define GEN_INPUT_FLUSH3(name, soft_t) \
static inline void name(soft_t *a, soft_t *b, soft_t *c, float_status *s) \
{ \
if (likely(!s->flush_inputs_to_zero)) { \
return; \
} \
soft_t ## _input_flush__nocheck(a, s); \
soft_t ## _input_flush__nocheck(b, s); \
soft_t ## _input_flush__nocheck(c, s); \
}
GEN_INPUT_FLUSH3(float32_input_flush3, float32)
GEN_INPUT_FLUSH3(float64_input_flush3, float64)
#undef GEN_INPUT_FLUSH3
/*
* Choose whether to use fpclassify or float32/64_* primitives in the generated
* hardfloat functions. Each combination of number of inputs and float size
* gets its own value.
*/
#if defined(__x86_64__)
# define QEMU_HARDFLOAT_1F32_USE_FP 0
# define QEMU_HARDFLOAT_1F64_USE_FP 1
# define QEMU_HARDFLOAT_2F32_USE_FP 0
# define QEMU_HARDFLOAT_2F64_USE_FP 1
# define QEMU_HARDFLOAT_3F32_USE_FP 0
# define QEMU_HARDFLOAT_3F64_USE_FP 1
#else
# define QEMU_HARDFLOAT_1F32_USE_FP 0
# define QEMU_HARDFLOAT_1F64_USE_FP 0
# define QEMU_HARDFLOAT_2F32_USE_FP 0
# define QEMU_HARDFLOAT_2F64_USE_FP 0
# define QEMU_HARDFLOAT_3F32_USE_FP 0
# define QEMU_HARDFLOAT_3F64_USE_FP 0
#endif
/*
* QEMU_HARDFLOAT_USE_ISINF chooses whether to use isinf() over
* float{32,64}_is_infinity when !USE_FP.
* On x86_64/aarch64, using the former over the latter can yield a ~6% speedup.
* On power64 however, using isinf() reduces fp-bench performance by up to 50%.
*/
#if defined(__x86_64__) || defined(__aarch64__)
# define QEMU_HARDFLOAT_USE_ISINF 1
#else
# define QEMU_HARDFLOAT_USE_ISINF 0
#endif
/*
* Some targets clear the FP flags before most FP operations. This prevents
* the use of hardfloat, since hardfloat relies on the inexact flag being
* already set.
*/
#if defined(TARGET_PPC) || defined(__FAST_MATH__)
# if defined(__FAST_MATH__)
# warning disabling hardfloat due to -ffast-math: hardfloat requires an exact \
IEEE implementation
# endif
# define QEMU_NO_HARDFLOAT 1
# define QEMU_SOFTFLOAT_ATTR QEMU_FLATTEN
#else
# define QEMU_NO_HARDFLOAT 0
# define QEMU_SOFTFLOAT_ATTR QEMU_FLATTEN __attribute__((noinline))
#endif
static inline bool can_use_fpu(const float_status *s)
{
if (QEMU_NO_HARDFLOAT) {
return false;
}
return likely(s->float_exception_flags & float_flag_inexact &&
s->float_rounding_mode == float_round_nearest_even);
}
/*
* Hardfloat generation functions. Each operation can have two flavors:
* either using softfloat primitives (e.g. float32_is_zero_or_normal) for
* most condition checks, or native ones (e.g. fpclassify).
*
* The flavor is chosen by the callers. Instead of using macros, we rely on the
* compiler to propagate constants and inline everything into the callers.
*
* We only generate functions for operations with two inputs, since only
* these are common enough to justify consolidating them into common code.
*/
typedef union {
float32 s;
float h;
} union_float32;
typedef union {
float64 s;
double h;
} union_float64;
typedef bool (*f32_check_fn)(union_float32 a, union_float32 b);
typedef bool (*f64_check_fn)(union_float64 a, union_float64 b);
typedef float32 (*soft_f32_op2_fn)(float32 a, float32 b, float_status *s);
typedef float64 (*soft_f64_op2_fn)(float64 a, float64 b, float_status *s);
typedef float (*hard_f32_op2_fn)(float a, float b);
typedef double (*hard_f64_op2_fn)(double a, double b);
/* 2-input is-zero-or-normal */
static inline bool f32_is_zon2(union_float32 a, union_float32 b)
{
if (QEMU_HARDFLOAT_2F32_USE_FP) {
/*
* Not using a temp variable for consecutive fpclassify calls ends up
* generating faster code.
*/
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
(fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO);
}
return float32_is_zero_or_normal(a.s) &&
float32_is_zero_or_normal(b.s);
}
static inline bool f64_is_zon2(union_float64 a, union_float64 b)
{
if (QEMU_HARDFLOAT_2F64_USE_FP) {
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
(fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO);
}
return float64_is_zero_or_normal(a.s) &&
float64_is_zero_or_normal(b.s);
}
/* 3-input is-zero-or-normal */
static inline
bool f32_is_zon3(union_float32 a, union_float32 b, union_float32 c)
{
if (QEMU_HARDFLOAT_3F32_USE_FP) {
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
(fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO) &&
(fpclassify(c.h) == FP_NORMAL || fpclassify(c.h) == FP_ZERO);
}
return float32_is_zero_or_normal(a.s) &&
float32_is_zero_or_normal(b.s) &&
float32_is_zero_or_normal(c.s);
}
static inline
bool f64_is_zon3(union_float64 a, union_float64 b, union_float64 c)
{
if (QEMU_HARDFLOAT_3F64_USE_FP) {
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
(fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO) &&
(fpclassify(c.h) == FP_NORMAL || fpclassify(c.h) == FP_ZERO);
}
return float64_is_zero_or_normal(a.s) &&
float64_is_zero_or_normal(b.s) &&
float64_is_zero_or_normal(c.s);
}
static inline bool f32_is_inf(union_float32 a)
{
if (QEMU_HARDFLOAT_USE_ISINF) {
return isinf(a.h);
}
return float32_is_infinity(a.s);
}
static inline bool f64_is_inf(union_float64 a)
{
if (QEMU_HARDFLOAT_USE_ISINF) {
return isinf(a.h);
}
return float64_is_infinity(a.s);
}
static inline float32
float32_gen2(float32 xa, float32 xb, float_status *s,
hard_f32_op2_fn hard, soft_f32_op2_fn soft,
f32_check_fn pre, f32_check_fn post)
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
{
union_float32 ua, ub, ur;
ua.s = xa;
ub.s = xb;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
float32_input_flush2(&ua.s, &ub.s, s);
if (unlikely(!pre(ua, ub))) {
goto soft;
}
ur.h = hard(ua.h, ub.h);
if (unlikely(f32_is_inf(ur))) {
float_raise(float_flag_overflow, s);
} else if (unlikely(fabsf(ur.h) <= FLT_MIN) && post(ua, ub)) {
goto soft;
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
}
return ur.s;
soft:
return soft(ua.s, ub.s, s);
}
static inline float64
float64_gen2(float64 xa, float64 xb, float_status *s,
hard_f64_op2_fn hard, soft_f64_op2_fn soft,
f64_check_fn pre, f64_check_fn post)
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
{
union_float64 ua, ub, ur;
ua.s = xa;
ub.s = xb;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
float64_input_flush2(&ua.s, &ub.s, s);
if (unlikely(!pre(ua, ub))) {
goto soft;
}
ur.h = hard(ua.h, ub.h);
if (unlikely(f64_is_inf(ur))) {
float_raise(float_flag_overflow, s);
} else if (unlikely(fabs(ur.h) <= DBL_MIN) && post(ua, ub)) {
goto soft;
fpu: introduce hardfloat The appended paves the way for leveraging the host FPU for a subset of guest FP operations. For most guest workloads (e.g. FP flags aren't ever cleared, inexact occurs often and rounding is set to the default [to nearest]) this will yield sizable performance speedups. The approach followed here avoids checking the FP exception flags register. See the added comment for details. This assumes that QEMU is running on an IEEE754-compliant FPU and that the rounding is set to the default (to nearest). The implementation-dependent specifics of the FPU should not matter; things like tininess detection and snan representation are still dealt with in soft-fp. However, this approach will break on most hosts if we compile QEMU with flags that break IEEE compatibility. There is no way to detect all of these flags at compilation time, but at least we check for -ffast-math (which defines __FAST_MATH__) and disable hardfloat (plus emit a #warning) when it is set. This patch just adds common code. Some operations will be migrated to hardfloat in subsequent patches to ease bisection. Note: some architectures (at least PPC, there might be others) clear the status flags passed to softfloat before most FP operations. This precludes the use of hardfloat, so to avoid introducing a performance regression for those targets, we add a flag to disable hardfloat. In the long run though it would be good to fix the targets so that at least the inexact flag passed to softfloat is indeed sticky. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:13:59 +03:00
}
return ur.s;
soft:
return soft(ua.s, ub.s, s);
}
/*
* Classify a floating point number. Everything above float_class_qnan
* is a NaN so cls >= float_class_qnan is any NaN.
*/
typedef enum __attribute__ ((__packed__)) {
float_class_unclassified,
float_class_zero,
float_class_normal,
float_class_inf,
float_class_qnan, /* all NaNs from here */
float_class_snan,
} FloatClass;
#define float_cmask(bit) (1u << (bit))
enum {
float_cmask_zero = float_cmask(float_class_zero),
float_cmask_normal = float_cmask(float_class_normal),
float_cmask_inf = float_cmask(float_class_inf),
float_cmask_qnan = float_cmask(float_class_qnan),
float_cmask_snan = float_cmask(float_class_snan),
float_cmask_infzero = float_cmask_zero | float_cmask_inf,
float_cmask_anynan = float_cmask_qnan | float_cmask_snan,
};
/* Flags for parts_minmax. */
enum {
/* Set for minimum; clear for maximum. */
minmax_ismin = 1,
/* Set for the IEEE 754-2008 minNum() and maxNum() operations. */
minmax_isnum = 2,
/* Set for the IEEE 754-2008 minNumMag() and minNumMag() operations. */
minmax_ismag = 4,
/*
* Set for the IEEE 754-2019 minimumNumber() and maximumNumber()
* operations.
*/
minmax_isnumber = 8,
};
/* Simple helpers for checking if, or what kind of, NaN we have */
static inline __attribute__((unused)) bool is_nan(FloatClass c)
{
return unlikely(c >= float_class_qnan);
}
static inline __attribute__((unused)) bool is_snan(FloatClass c)
{
return c == float_class_snan;
}
static inline __attribute__((unused)) bool is_qnan(FloatClass c)
{
return c == float_class_qnan;
}
/*
* Structure holding all of the decomposed parts of a float.
* The exponent is unbiased and the fraction is normalized.
*
* The fraction words are stored in big-endian word ordering,
* so that truncation from a larger format to a smaller format
* can be done simply by ignoring subsequent elements.
*/
typedef struct {
FloatClass cls;
bool sign;
int32_t exp;
union {
/* Routines that know the structure may reference the singular name. */
uint64_t frac;
/*
* Routines expanded with multiple structures reference "hi" and "lo"
* depending on the operation. In FloatParts64, "hi" and "lo" are
* both the same word and aliased here.
*/
uint64_t frac_hi;
uint64_t frac_lo;
};
} FloatParts64;
typedef struct {
FloatClass cls;
bool sign;
int32_t exp;
uint64_t frac_hi;
uint64_t frac_lo;
} FloatParts128;
typedef struct {
FloatClass cls;
bool sign;
int32_t exp;
uint64_t frac_hi;
uint64_t frac_hm; /* high-middle */
uint64_t frac_lm; /* low-middle */
uint64_t frac_lo;
} FloatParts256;
/* These apply to the most significant word of each FloatPartsN. */
#define DECOMPOSED_BINARY_POINT 63
#define DECOMPOSED_IMPLICIT_BIT (1ull << DECOMPOSED_BINARY_POINT)
/* Structure holding all of the relevant parameters for a format.
* exp_size: the size of the exponent field
* exp_bias: the offset applied to the exponent field
* exp_max: the maximum normalised exponent
* frac_size: the size of the fraction field
* frac_shift: shift to normalise the fraction with DECOMPOSED_BINARY_POINT
* The following are computed based the size of fraction
* round_mask: bits below lsb which must be rounded
* The following optional modifiers are available:
* arm_althp: handle ARM Alternative Half Precision
*/
typedef struct {
int exp_size;
int exp_bias;
int exp_re_bias;
int exp_max;
int frac_size;
int frac_shift;
bool arm_althp;
uint64_t round_mask;
} FloatFmt;
/* Expand fields based on the size of exponent and fraction */
#define FLOAT_PARAMS_(E) \
.exp_size = E, \
.exp_bias = ((1 << E) - 1) >> 1, \
.exp_re_bias = (1 << (E - 1)) + (1 << (E - 2)), \
.exp_max = (1 << E) - 1
#define FLOAT_PARAMS(E, F) \
FLOAT_PARAMS_(E), \
.frac_size = F, \
.frac_shift = (-F - 1) & 63, \
.round_mask = (1ull << ((-F - 1) & 63)) - 1
static const FloatFmt float16_params = {
FLOAT_PARAMS(5, 10)
};
static const FloatFmt float16_params_ahp = {
FLOAT_PARAMS(5, 10),
.arm_althp = true
};
static const FloatFmt bfloat16_params = {
FLOAT_PARAMS(8, 7)
};
static const FloatFmt float32_params = {
FLOAT_PARAMS(8, 23)
};
static const FloatFmt float64_params = {
FLOAT_PARAMS(11, 52)
};
static const FloatFmt float128_params = {
FLOAT_PARAMS(15, 112)
};
#define FLOATX80_PARAMS(R) \
FLOAT_PARAMS_(15), \
.frac_size = R == 64 ? 63 : R, \
.frac_shift = 0, \
.round_mask = R == 64 ? -1 : (1ull << ((-R - 1) & 63)) - 1
static const FloatFmt floatx80_params[3] = {
[floatx80_precision_s] = { FLOATX80_PARAMS(23) },
[floatx80_precision_d] = { FLOATX80_PARAMS(52) },
[floatx80_precision_x] = { FLOATX80_PARAMS(64) },
};
/* Unpack a float to parts, but do not canonicalize. */
static void unpack_raw64(FloatParts64 *r, const FloatFmt *fmt, uint64_t raw)
{
const int f_size = fmt->frac_size;
const int e_size = fmt->exp_size;
*r = (FloatParts64) {
.cls = float_class_unclassified,
.sign = extract64(raw, f_size + e_size, 1),
.exp = extract64(raw, f_size, e_size),
.frac = extract64(raw, 0, f_size)
};
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN float16_unpack_raw(FloatParts64 *p, float16 f)
{
unpack_raw64(p, &float16_params, f);
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN bfloat16_unpack_raw(FloatParts64 *p, bfloat16 f)
{
unpack_raw64(p, &bfloat16_params, f);
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN float32_unpack_raw(FloatParts64 *p, float32 f)
{
unpack_raw64(p, &float32_params, f);
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN float64_unpack_raw(FloatParts64 *p, float64 f)
{
unpack_raw64(p, &float64_params, f);
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN floatx80_unpack_raw(FloatParts128 *p, floatx80 f)
{
*p = (FloatParts128) {
.cls = float_class_unclassified,
.sign = extract32(f.high, 15, 1),
.exp = extract32(f.high, 0, 15),
.frac_hi = f.low
};
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static void QEMU_FLATTEN float128_unpack_raw(FloatParts128 *p, float128 f)
{
const int f_size = float128_params.frac_size - 64;
const int e_size = float128_params.exp_size;
*p = (FloatParts128) {
.cls = float_class_unclassified,
.sign = extract64(f.high, f_size + e_size, 1),
.exp = extract64(f.high, f_size, e_size),
.frac_hi = extract64(f.high, 0, f_size),
.frac_lo = f.low,
};
}
/* Pack a float from parts, but do not canonicalize. */
static uint64_t pack_raw64(const FloatParts64 *p, const FloatFmt *fmt)
{
const int f_size = fmt->frac_size;
const int e_size = fmt->exp_size;
uint64_t ret;
ret = (uint64_t)p->sign << (f_size + e_size);
ret = deposit64(ret, f_size, e_size, p->exp);
ret = deposit64(ret, 0, f_size, p->frac);
return ret;
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static float16 QEMU_FLATTEN float16_pack_raw(const FloatParts64 *p)
{
return make_float16(pack_raw64(p, &float16_params));
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static bfloat16 QEMU_FLATTEN bfloat16_pack_raw(const FloatParts64 *p)
{
return pack_raw64(p, &bfloat16_params);
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static float32 QEMU_FLATTEN float32_pack_raw(const FloatParts64 *p)
{
return make_float32(pack_raw64(p, &float32_params));
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static float64 QEMU_FLATTEN float64_pack_raw(const FloatParts64 *p)
{
return make_float64(pack_raw64(p, &float64_params));
}
softfloat: use QEMU_FLATTEN to avoid mistaken isra inlining Balton discovered that asserts for the extract/deposit calls had a significant impact on a lame benchmark on qemu-ppc. Replicating with: ./qemu-ppc64 ~/lsrc/tests/lame.git-svn/builds/ppc64/frontend/lame \ -h pts-trondheim-3.wav pts-trondheim-3.mp3 showed up the pack/unpack routines not eliding the assert checks as it should have done causing them to prominently figure in the profile:  11.44%  qemu-ppc64  qemu-ppc64               [.] unpack_raw64.isra.0  11.03%  qemu-ppc64  qemu-ppc64               [.] parts64_uncanon_normal   8.26%  qemu-ppc64  qemu-ppc64               [.] helper_compute_fprf_float64   6.75%  qemu-ppc64  qemu-ppc64               [.] do_float_check_status   5.34%  qemu-ppc64  qemu-ppc64               [.] parts64_muladd   4.75%  qemu-ppc64  qemu-ppc64               [.] pack_raw64.isra.0   4.38%  qemu-ppc64  qemu-ppc64               [.] parts64_canonicalize   3.62%  qemu-ppc64  qemu-ppc64               [.] float64r32_round_pack_canonical After this patch the same test runs 31 seconds faster with a profile where the generated code dominates more: + 14.12% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619420 + 13.30% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000616850 + 12.58% 12.19% qemu-ppc64 qemu-ppc64 [.] parts64_uncanon_normal + 10.62% 0.00% qemu-ppc64 [unknown] [.] 0x000000400061bf70 + 9.91% 9.73% qemu-ppc64 qemu-ppc64 [.] helper_compute_fprf_float64 + 7.84% 7.82% qemu-ppc64 qemu-ppc64 [.] do_float_check_status + 6.47% 5.78% qemu-ppc64 qemu-ppc64 [.] parts64_canonicalize.constprop.0 + 6.46% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000620130 + 6.42% 0.00% qemu-ppc64 [unknown] [.] 0x0000004000619400 + 6.17% 6.04% qemu-ppc64 qemu-ppc64 [.] parts64_muladd + 5.85% 0.00% qemu-ppc64 [unknown] [.] 0x00000040006167e0 + 5.74% 0.00% qemu-ppc64 [unknown] [.] 0x0000b693fcffffd3 + 5.45% 4.78% qemu-ppc64 qemu-ppc64 [.] float64r32_round_pack_canonical Suggested-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <ec9cfe5a-d5f2-466d-34dc-c35817e7e010@linaro.org> [AJB: Patchified rth's suggestion] Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Cc: BALATON Zoltan <balaton@eik.bme.hu> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Tested-by: BALATON Zoltan <balaton@eik.bme.hu> Message-Id: <20230523131107.3680641-1-alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2023-05-23 16:11:07 +03:00
static float128 QEMU_FLATTEN float128_pack_raw(const FloatParts128 *p)
{
const int f_size = float128_params.frac_size - 64;
const int e_size = float128_params.exp_size;
uint64_t hi;
hi = (uint64_t)p->sign << (f_size + e_size);
hi = deposit64(hi, f_size, e_size, p->exp);
hi = deposit64(hi, 0, f_size, p->frac_hi);
return make_float128(hi, p->frac_lo);
}
/*----------------------------------------------------------------------------
| Functions and definitions to determine: (1) whether tininess for underflow
| is detected before or after rounding by default, (2) what (if anything)
| happens when exceptions are raised, (3) how signaling NaNs are distinguished
| from quiet NaNs, (4) the default generated quiet NaNs, and (5) how NaNs
| are propagated from function inputs to output. These details are target-
| specific.
*----------------------------------------------------------------------------*/
#include "softfloat-specialize.c.inc"
#define PARTS_GENERIC_64_128(NAME, P) \
_Generic((P), FloatParts64 *: parts64_##NAME, \
FloatParts128 *: parts128_##NAME)
#define PARTS_GENERIC_64_128_256(NAME, P) \
_Generic((P), FloatParts64 *: parts64_##NAME, \
FloatParts128 *: parts128_##NAME, \
FloatParts256 *: parts256_##NAME)
#define parts_default_nan(P, S) PARTS_GENERIC_64_128(default_nan, P)(P, S)
#define parts_silence_nan(P, S) PARTS_GENERIC_64_128(silence_nan, P)(P, S)
static void parts64_return_nan(FloatParts64 *a, float_status *s);
static void parts128_return_nan(FloatParts128 *a, float_status *s);
#define parts_return_nan(P, S) PARTS_GENERIC_64_128(return_nan, P)(P, S)
static FloatParts64 *parts64_pick_nan(FloatParts64 *a, FloatParts64 *b,
float_status *s);
static FloatParts128 *parts128_pick_nan(FloatParts128 *a, FloatParts128 *b,
float_status *s);
#define parts_pick_nan(A, B, S) PARTS_GENERIC_64_128(pick_nan, A)(A, B, S)
static FloatParts64 *parts64_pick_nan_muladd(FloatParts64 *a, FloatParts64 *b,
FloatParts64 *c, float_status *s,
int ab_mask, int abc_mask);
static FloatParts128 *parts128_pick_nan_muladd(FloatParts128 *a,
FloatParts128 *b,
FloatParts128 *c,
float_status *s,
int ab_mask, int abc_mask);
#define parts_pick_nan_muladd(A, B, C, S, ABM, ABCM) \
PARTS_GENERIC_64_128(pick_nan_muladd, A)(A, B, C, S, ABM, ABCM)
static void parts64_canonicalize(FloatParts64 *p, float_status *status,
const FloatFmt *fmt);
static void parts128_canonicalize(FloatParts128 *p, float_status *status,
const FloatFmt *fmt);
#define parts_canonicalize(A, S, F) \
PARTS_GENERIC_64_128(canonicalize, A)(A, S, F)
static void parts64_uncanon_normal(FloatParts64 *p, float_status *status,
const FloatFmt *fmt);
static void parts128_uncanon_normal(FloatParts128 *p, float_status *status,
const FloatFmt *fmt);
#define parts_uncanon_normal(A, S, F) \
PARTS_GENERIC_64_128(uncanon_normal, A)(A, S, F)
static void parts64_uncanon(FloatParts64 *p, float_status *status,
const FloatFmt *fmt);
static void parts128_uncanon(FloatParts128 *p, float_status *status,
const FloatFmt *fmt);
#define parts_uncanon(A, S, F) \
PARTS_GENERIC_64_128(uncanon, A)(A, S, F)
static void parts64_add_normal(FloatParts64 *a, FloatParts64 *b);
static void parts128_add_normal(FloatParts128 *a, FloatParts128 *b);
static void parts256_add_normal(FloatParts256 *a, FloatParts256 *b);
#define parts_add_normal(A, B) \
PARTS_GENERIC_64_128_256(add_normal, A)(A, B)
static bool parts64_sub_normal(FloatParts64 *a, FloatParts64 *b);
static bool parts128_sub_normal(FloatParts128 *a, FloatParts128 *b);
static bool parts256_sub_normal(FloatParts256 *a, FloatParts256 *b);
#define parts_sub_normal(A, B) \
PARTS_GENERIC_64_128_256(sub_normal, A)(A, B)
static FloatParts64 *parts64_addsub(FloatParts64 *a, FloatParts64 *b,
float_status *s, bool subtract);
static FloatParts128 *parts128_addsub(FloatParts128 *a, FloatParts128 *b,
float_status *s, bool subtract);
#define parts_addsub(A, B, S, Z) \
PARTS_GENERIC_64_128(addsub, A)(A, B, S, Z)
static FloatParts64 *parts64_mul(FloatParts64 *a, FloatParts64 *b,
float_status *s);
static FloatParts128 *parts128_mul(FloatParts128 *a, FloatParts128 *b,
float_status *s);
#define parts_mul(A, B, S) \
PARTS_GENERIC_64_128(mul, A)(A, B, S)
static FloatParts64 *parts64_muladd(FloatParts64 *a, FloatParts64 *b,
FloatParts64 *c, int flags,
float_status *s);
static FloatParts128 *parts128_muladd(FloatParts128 *a, FloatParts128 *b,
FloatParts128 *c, int flags,
float_status *s);
#define parts_muladd(A, B, C, Z, S) \
PARTS_GENERIC_64_128(muladd, A)(A, B, C, Z, S)
static FloatParts64 *parts64_div(FloatParts64 *a, FloatParts64 *b,
float_status *s);
static FloatParts128 *parts128_div(FloatParts128 *a, FloatParts128 *b,
float_status *s);
#define parts_div(A, B, S) \
PARTS_GENERIC_64_128(div, A)(A, B, S)
static FloatParts64 *parts64_modrem(FloatParts64 *a, FloatParts64 *b,
uint64_t *mod_quot, float_status *s);
static FloatParts128 *parts128_modrem(FloatParts128 *a, FloatParts128 *b,
uint64_t *mod_quot, float_status *s);
#define parts_modrem(A, B, Q, S) \
PARTS_GENERIC_64_128(modrem, A)(A, B, Q, S)
static void parts64_sqrt(FloatParts64 *a, float_status *s, const FloatFmt *f);
static void parts128_sqrt(FloatParts128 *a, float_status *s, const FloatFmt *f);
#define parts_sqrt(A, S, F) \
PARTS_GENERIC_64_128(sqrt, A)(A, S, F)
static bool parts64_round_to_int_normal(FloatParts64 *a, FloatRoundMode rm,
int scale, int frac_size);
static bool parts128_round_to_int_normal(FloatParts128 *a, FloatRoundMode r,
int scale, int frac_size);
#define parts_round_to_int_normal(A, R, C, F) \
PARTS_GENERIC_64_128(round_to_int_normal, A)(A, R, C, F)
static void parts64_round_to_int(FloatParts64 *a, FloatRoundMode rm,
int scale, float_status *s,
const FloatFmt *fmt);
static void parts128_round_to_int(FloatParts128 *a, FloatRoundMode r,
int scale, float_status *s,
const FloatFmt *fmt);
#define parts_round_to_int(A, R, C, S, F) \
PARTS_GENERIC_64_128(round_to_int, A)(A, R, C, S, F)
static int64_t parts64_float_to_sint(FloatParts64 *p, FloatRoundMode rmode,
int scale, int64_t min, int64_t max,
float_status *s);
static int64_t parts128_float_to_sint(FloatParts128 *p, FloatRoundMode rmode,
int scale, int64_t min, int64_t max,
float_status *s);
#define parts_float_to_sint(P, R, Z, MN, MX, S) \
PARTS_GENERIC_64_128(float_to_sint, P)(P, R, Z, MN, MX, S)
static uint64_t parts64_float_to_uint(FloatParts64 *p, FloatRoundMode rmode,
int scale, uint64_t max,
float_status *s);
static uint64_t parts128_float_to_uint(FloatParts128 *p, FloatRoundMode rmode,
int scale, uint64_t max,
float_status *s);
#define parts_float_to_uint(P, R, Z, M, S) \
PARTS_GENERIC_64_128(float_to_uint, P)(P, R, Z, M, S)
static int64_t parts64_float_to_sint_modulo(FloatParts64 *p,
FloatRoundMode rmode,
int bitsm1, float_status *s);
static int64_t parts128_float_to_sint_modulo(FloatParts128 *p,
FloatRoundMode rmode,
int bitsm1, float_status *s);
#define parts_float_to_sint_modulo(P, R, M, S) \
PARTS_GENERIC_64_128(float_to_sint_modulo, P)(P, R, M, S)
static void parts64_sint_to_float(FloatParts64 *p, int64_t a,
int scale, float_status *s);
static void parts128_sint_to_float(FloatParts128 *p, int64_t a,
int scale, float_status *s);
#define parts_float_to_sint(P, R, Z, MN, MX, S) \
PARTS_GENERIC_64_128(float_to_sint, P)(P, R, Z, MN, MX, S)
#define parts_sint_to_float(P, I, Z, S) \
PARTS_GENERIC_64_128(sint_to_float, P)(P, I, Z, S)
static void parts64_uint_to_float(FloatParts64 *p, uint64_t a,
int scale, float_status *s);
static void parts128_uint_to_float(FloatParts128 *p, uint64_t a,
int scale, float_status *s);
#define parts_uint_to_float(P, I, Z, S) \
PARTS_GENERIC_64_128(uint_to_float, P)(P, I, Z, S)
static FloatParts64 *parts64_minmax(FloatParts64 *a, FloatParts64 *b,
float_status *s, int flags);
static FloatParts128 *parts128_minmax(FloatParts128 *a, FloatParts128 *b,
float_status *s, int flags);
#define parts_minmax(A, B, S, F) \
PARTS_GENERIC_64_128(minmax, A)(A, B, S, F)
static FloatRelation parts64_compare(FloatParts64 *a, FloatParts64 *b,
float_status *s, bool q);
static FloatRelation parts128_compare(FloatParts128 *a, FloatParts128 *b,
float_status *s, bool q);
#define parts_compare(A, B, S, Q) \
PARTS_GENERIC_64_128(compare, A)(A, B, S, Q)
static void parts64_scalbn(FloatParts64 *a, int n, float_status *s);
static void parts128_scalbn(FloatParts128 *a, int n, float_status *s);
#define parts_scalbn(A, N, S) \
PARTS_GENERIC_64_128(scalbn, A)(A, N, S)
static void parts64_log2(FloatParts64 *a, float_status *s, const FloatFmt *f);
static void parts128_log2(FloatParts128 *a, float_status *s, const FloatFmt *f);
#define parts_log2(A, S, F) \
PARTS_GENERIC_64_128(log2, A)(A, S, F)
/*
* Helper functions for softfloat-parts.c.inc, per-size operations.
*/
#define FRAC_GENERIC_64_128(NAME, P) \
_Generic((P), FloatParts64 *: frac64_##NAME, \
FloatParts128 *: frac128_##NAME)
#define FRAC_GENERIC_64_128_256(NAME, P) \
_Generic((P), FloatParts64 *: frac64_##NAME, \
FloatParts128 *: frac128_##NAME, \
FloatParts256 *: frac256_##NAME)
static bool frac64_add(FloatParts64 *r, FloatParts64 *a, FloatParts64 *b)
{
return uadd64_overflow(a->frac, b->frac, &r->frac);
}
static bool frac128_add(FloatParts128 *r, FloatParts128 *a, FloatParts128 *b)
{
bool c = 0;
r->frac_lo = uadd64_carry(a->frac_lo, b->frac_lo, &c);
r->frac_hi = uadd64_carry(a->frac_hi, b->frac_hi, &c);
return c;
}
static bool frac256_add(FloatParts256 *r, FloatParts256 *a, FloatParts256 *b)
{
bool c = 0;
r->frac_lo = uadd64_carry(a->frac_lo, b->frac_lo, &c);
r->frac_lm = uadd64_carry(a->frac_lm, b->frac_lm, &c);
r->frac_hm = uadd64_carry(a->frac_hm, b->frac_hm, &c);
r->frac_hi = uadd64_carry(a->frac_hi, b->frac_hi, &c);
return c;
}
#define frac_add(R, A, B) FRAC_GENERIC_64_128_256(add, R)(R, A, B)
static bool frac64_addi(FloatParts64 *r, FloatParts64 *a, uint64_t c)
{
return uadd64_overflow(a->frac, c, &r->frac);
}
static bool frac128_addi(FloatParts128 *r, FloatParts128 *a, uint64_t c)
{
c = uadd64_overflow(a->frac_lo, c, &r->frac_lo);
return uadd64_overflow(a->frac_hi, c, &r->frac_hi);
}
#define frac_addi(R, A, C) FRAC_GENERIC_64_128(addi, R)(R, A, C)
static void frac64_allones(FloatParts64 *a)
{
a->frac = -1;
}
static void frac128_allones(FloatParts128 *a)
{
a->frac_hi = a->frac_lo = -1;
}
#define frac_allones(A) FRAC_GENERIC_64_128(allones, A)(A)
static FloatRelation frac64_cmp(FloatParts64 *a, FloatParts64 *b)
{
return (a->frac == b->frac ? float_relation_equal
: a->frac < b->frac ? float_relation_less
: float_relation_greater);
}
static FloatRelation frac128_cmp(FloatParts128 *a, FloatParts128 *b)
{
uint64_t ta = a->frac_hi, tb = b->frac_hi;
if (ta == tb) {
ta = a->frac_lo, tb = b->frac_lo;
if (ta == tb) {
return float_relation_equal;
}
}
return ta < tb ? float_relation_less : float_relation_greater;
}
#define frac_cmp(A, B) FRAC_GENERIC_64_128(cmp, A)(A, B)
static void frac64_clear(FloatParts64 *a)
{
a->frac = 0;
}
static void frac128_clear(FloatParts128 *a)
{
a->frac_hi = a->frac_lo = 0;
}
#define frac_clear(A) FRAC_GENERIC_64_128(clear, A)(A)
static bool frac64_div(FloatParts64 *a, FloatParts64 *b)
{
uint64_t n1, n0, r, q;
bool ret;
/*
* We want a 2*N / N-bit division to produce exactly an N-bit
* result, so that we do not lose any precision and so that we
* do not have to renormalize afterward. If A.frac < B.frac,
* then division would produce an (N-1)-bit result; shift A left
* by one to produce the an N-bit result, and return true to
* decrement the exponent to match.
*
* The udiv_qrnnd algorithm that we're using requires normalization,
* i.e. the msb of the denominator must be set, which is already true.
*/
ret = a->frac < b->frac;
if (ret) {
n0 = a->frac;
n1 = 0;
} else {
n0 = a->frac >> 1;
n1 = a->frac << 63;
}
q = udiv_qrnnd(&r, n0, n1, b->frac);
/* Set lsb if there is a remainder, to set inexact. */
a->frac = q | (r != 0);
return ret;
}
static bool frac128_div(FloatParts128 *a, FloatParts128 *b)
{
uint64_t q0, q1, a0, a1, b0, b1;
uint64_t r0, r1, r2, r3, t0, t1, t2, t3;
bool ret = false;
a0 = a->frac_hi, a1 = a->frac_lo;
b0 = b->frac_hi, b1 = b->frac_lo;
ret = lt128(a0, a1, b0, b1);
if (!ret) {
a1 = shr_double(a0, a1, 1);
a0 = a0 >> 1;
}
/* Use 128/64 -> 64 division as estimate for 192/128 -> 128 division. */
q0 = estimateDiv128To64(a0, a1, b0);
/*
* Estimate is high because B1 was not included (unless B1 == 0).
* Reduce quotient and increase remainder until remainder is non-negative.
* This loop will execute 0 to 2 times.
*/
mul128By64To192(b0, b1, q0, &t0, &t1, &t2);
sub192(a0, a1, 0, t0, t1, t2, &r0, &r1, &r2);
while (r0 != 0) {
q0--;
add192(r0, r1, r2, 0, b0, b1, &r0, &r1, &r2);
}
/* Repeat using the remainder, producing a second word of quotient. */
q1 = estimateDiv128To64(r1, r2, b0);
mul128By64To192(b0, b1, q1, &t1, &t2, &t3);
sub192(r1, r2, 0, t1, t2, t3, &r1, &r2, &r3);
while (r1 != 0) {
q1--;
add192(r1, r2, r3, 0, b0, b1, &r1, &r2, &r3);
}
/* Any remainder indicates inexact; set sticky bit. */
q1 |= (r2 | r3) != 0;
a->frac_hi = q0;
a->frac_lo = q1;
return ret;
}
#define frac_div(A, B) FRAC_GENERIC_64_128(div, A)(A, B)
static bool frac64_eqz(FloatParts64 *a)
{
return a->frac == 0;
}
static bool frac128_eqz(FloatParts128 *a)
{
return (a->frac_hi | a->frac_lo) == 0;
}
#define frac_eqz(A) FRAC_GENERIC_64_128(eqz, A)(A)
static void frac64_mulw(FloatParts128 *r, FloatParts64 *a, FloatParts64 *b)
{
mulu64(&r->frac_lo, &r->frac_hi, a->frac, b->frac);
}
static void frac128_mulw(FloatParts256 *r, FloatParts128 *a, FloatParts128 *b)
{
mul128To256(a->frac_hi, a->frac_lo, b->frac_hi, b->frac_lo,
&r->frac_hi, &r->frac_hm, &r->frac_lm, &r->frac_lo);
}
#define frac_mulw(R, A, B) FRAC_GENERIC_64_128(mulw, A)(R, A, B)
static void frac64_neg(FloatParts64 *a)
{
a->frac = -a->frac;
}
static void frac128_neg(FloatParts128 *a)
{
bool c = 0;
a->frac_lo = usub64_borrow(0, a->frac_lo, &c);
a->frac_hi = usub64_borrow(0, a->frac_hi, &c);
}
static void frac256_neg(FloatParts256 *a)
{
bool c = 0;
a->frac_lo = usub64_borrow(0, a->frac_lo, &c);
a->frac_lm = usub64_borrow(0, a->frac_lm, &c);
a->frac_hm = usub64_borrow(0, a->frac_hm, &c);
a->frac_hi = usub64_borrow(0, a->frac_hi, &c);
}
#define frac_neg(A) FRAC_GENERIC_64_128_256(neg, A)(A)
static int frac64_normalize(FloatParts64 *a)
{
if (a->frac) {
int shift = clz64(a->frac);
a->frac <<= shift;
return shift;
}
return 64;
}
static int frac128_normalize(FloatParts128 *a)
{
if (a->frac_hi) {
int shl = clz64(a->frac_hi);
a->frac_hi = shl_double(a->frac_hi, a->frac_lo, shl);
a->frac_lo <<= shl;
return shl;
} else if (a->frac_lo) {
int shl = clz64(a->frac_lo);
a->frac_hi = a->frac_lo << shl;
a->frac_lo = 0;
return shl + 64;
}
return 128;
}
static int frac256_normalize(FloatParts256 *a)
{
uint64_t a0 = a->frac_hi, a1 = a->frac_hm;
uint64_t a2 = a->frac_lm, a3 = a->frac_lo;
int ret, shl;
if (likely(a0)) {
shl = clz64(a0);
if (shl == 0) {
return 0;
}
ret = shl;
} else {
if (a1) {
ret = 64;
a0 = a1, a1 = a2, a2 = a3, a3 = 0;
} else if (a2) {
ret = 128;
a0 = a2, a1 = a3, a2 = 0, a3 = 0;
} else if (a3) {
ret = 192;
a0 = a3, a1 = 0, a2 = 0, a3 = 0;
} else {
ret = 256;
a0 = 0, a1 = 0, a2 = 0, a3 = 0;
goto done;
}
shl = clz64(a0);
if (shl == 0) {
goto done;
}
ret += shl;
}
a0 = shl_double(a0, a1, shl);
a1 = shl_double(a1, a2, shl);
a2 = shl_double(a2, a3, shl);
a3 <<= shl;
done:
a->frac_hi = a0;
a->frac_hm = a1;
a->frac_lm = a2;
a->frac_lo = a3;
return ret;
}
#define frac_normalize(A) FRAC_GENERIC_64_128_256(normalize, A)(A)
static void frac64_modrem(FloatParts64 *a, FloatParts64 *b, uint64_t *mod_quot)
{
uint64_t a0, a1, b0, t0, t1, q, quot;
int exp_diff = a->exp - b->exp;
int shift;
a0 = a->frac;
a1 = 0;
if (exp_diff < -1) {
if (mod_quot) {
*mod_quot = 0;
}
return;
}
if (exp_diff == -1) {
a0 >>= 1;
exp_diff = 0;
}
b0 = b->frac;
quot = q = b0 <= a0;
if (q) {
a0 -= b0;
}
exp_diff -= 64;
while (exp_diff > 0) {
q = estimateDiv128To64(a0, a1, b0);
q = q > 2 ? q - 2 : 0;
mul64To128(b0, q, &t0, &t1);
sub128(a0, a1, t0, t1, &a0, &a1);
shortShift128Left(a0, a1, 62, &a0, &a1);
exp_diff -= 62;
quot = (quot << 62) + q;
}
exp_diff += 64;
if (exp_diff > 0) {
q = estimateDiv128To64(a0, a1, b0);
q = q > 2 ? (q - 2) >> (64 - exp_diff) : 0;
mul64To128(b0, q << (64 - exp_diff), &t0, &t1);
sub128(a0, a1, t0, t1, &a0, &a1);
shortShift128Left(0, b0, 64 - exp_diff, &t0, &t1);
while (le128(t0, t1, a0, a1)) {
++q;
sub128(a0, a1, t0, t1, &a0, &a1);
}
quot = (exp_diff < 64 ? quot << exp_diff : 0) + q;
} else {
t0 = b0;
t1 = 0;
}
if (mod_quot) {
*mod_quot = quot;
} else {
sub128(t0, t1, a0, a1, &t0, &t1);
if (lt128(t0, t1, a0, a1) ||
(eq128(t0, t1, a0, a1) && (q & 1))) {
a0 = t0;
a1 = t1;
a->sign = !a->sign;
}
}
if (likely(a0)) {
shift = clz64(a0);
shortShift128Left(a0, a1, shift, &a0, &a1);
} else if (likely(a1)) {
shift = clz64(a1);
a0 = a1 << shift;
a1 = 0;
shift += 64;
} else {
a->cls = float_class_zero;
return;
}
a->exp = b->exp + exp_diff - shift;
a->frac = a0 | (a1 != 0);
}
static void frac128_modrem(FloatParts128 *a, FloatParts128 *b,
uint64_t *mod_quot)
{
uint64_t a0, a1, a2, b0, b1, t0, t1, t2, q, quot;
int exp_diff = a->exp - b->exp;
int shift;
a0 = a->frac_hi;
a1 = a->frac_lo;
a2 = 0;
if (exp_diff < -1) {
if (mod_quot) {
*mod_quot = 0;
}
return;
}
if (exp_diff == -1) {
shift128Right(a0, a1, 1, &a0, &a1);
exp_diff = 0;
}
b0 = b->frac_hi;
b1 = b->frac_lo;
quot = q = le128(b0, b1, a0, a1);
if (q) {
sub128(a0, a1, b0, b1, &a0, &a1);
}
exp_diff -= 64;
while (exp_diff > 0) {
q = estimateDiv128To64(a0, a1, b0);
q = q > 4 ? q - 4 : 0;
mul128By64To192(b0, b1, q, &t0, &t1, &t2);
sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2);
shortShift192Left(a0, a1, a2, 61, &a0, &a1, &a2);
exp_diff -= 61;
quot = (quot << 61) + q;
}
exp_diff += 64;
if (exp_diff > 0) {
q = estimateDiv128To64(a0, a1, b0);
q = q > 4 ? (q - 4) >> (64 - exp_diff) : 0;
mul128By64To192(b0, b1, q << (64 - exp_diff), &t0, &t1, &t2);
sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2);
shortShift192Left(0, b0, b1, 64 - exp_diff, &t0, &t1, &t2);
while (le192(t0, t1, t2, a0, a1, a2)) {
++q;
sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2);
}
quot = (exp_diff < 64 ? quot << exp_diff : 0) + q;
} else {
t0 = b0;
t1 = b1;
t2 = 0;
}
if (mod_quot) {
*mod_quot = quot;
} else {
sub192(t0, t1, t2, a0, a1, a2, &t0, &t1, &t2);
if (lt192(t0, t1, t2, a0, a1, a2) ||
(eq192(t0, t1, t2, a0, a1, a2) && (q & 1))) {
a0 = t0;
a1 = t1;
a2 = t2;
a->sign = !a->sign;
}
}
if (likely(a0)) {
shift = clz64(a0);
shortShift192Left(a0, a1, a2, shift, &a0, &a1, &a2);
} else if (likely(a1)) {
shift = clz64(a1);
shortShift128Left(a1, a2, shift, &a0, &a1);
a2 = 0;
shift += 64;
} else if (likely(a2)) {
shift = clz64(a2);
a0 = a2 << shift;
a1 = a2 = 0;
shift += 128;
} else {
a->cls = float_class_zero;
return;
}
a->exp = b->exp + exp_diff - shift;
a->frac_hi = a0;
a->frac_lo = a1 | (a2 != 0);
}
#define frac_modrem(A, B, Q) FRAC_GENERIC_64_128(modrem, A)(A, B, Q)
static void frac64_shl(FloatParts64 *a, int c)
{
a->frac <<= c;
}
static void frac128_shl(FloatParts128 *a, int c)
{
uint64_t a0 = a->frac_hi, a1 = a->frac_lo;
if (c & 64) {
a0 = a1, a1 = 0;
}
c &= 63;
if (c) {
a0 = shl_double(a0, a1, c);
a1 = a1 << c;
}
a->frac_hi = a0;
a->frac_lo = a1;
}
#define frac_shl(A, C) FRAC_GENERIC_64_128(shl, A)(A, C)
static void frac64_shr(FloatParts64 *a, int c)
{
a->frac >>= c;
}
static void frac128_shr(FloatParts128 *a, int c)
{
uint64_t a0 = a->frac_hi, a1 = a->frac_lo;
if (c & 64) {
a1 = a0, a0 = 0;
}
c &= 63;
if (c) {
a1 = shr_double(a0, a1, c);
a0 = a0 >> c;
}
a->frac_hi = a0;
a->frac_lo = a1;
}
#define frac_shr(A, C) FRAC_GENERIC_64_128(shr, A)(A, C)
static void frac64_shrjam(FloatParts64 *a, int c)
{
uint64_t a0 = a->frac;
if (likely(c != 0)) {
if (likely(c < 64)) {
a0 = (a0 >> c) | (shr_double(a0, 0, c) != 0);
} else {
a0 = a0 != 0;
}
a->frac = a0;
}
}
static void frac128_shrjam(FloatParts128 *a, int c)
{
uint64_t a0 = a->frac_hi, a1 = a->frac_lo;
uint64_t sticky = 0;
if (unlikely(c == 0)) {
return;
} else if (likely(c < 64)) {
/* nothing */
} else if (likely(c < 128)) {
sticky = a1;
a1 = a0;
a0 = 0;
c &= 63;
if (c == 0) {
goto done;
}
} else {
sticky = a0 | a1;
a0 = a1 = 0;
goto done;
}
sticky |= shr_double(a1, 0, c);
a1 = shr_double(a0, a1, c);
a0 = a0 >> c;
done:
a->frac_lo = a1 | (sticky != 0);
a->frac_hi = a0;
}
static void frac256_shrjam(FloatParts256 *a, int c)
{
uint64_t a0 = a->frac_hi, a1 = a->frac_hm;
uint64_t a2 = a->frac_lm, a3 = a->frac_lo;
uint64_t sticky = 0;
if (unlikely(c == 0)) {
return;
} else if (likely(c < 64)) {
/* nothing */
} else if (likely(c < 256)) {
if (unlikely(c & 128)) {
sticky |= a2 | a3;
a3 = a1, a2 = a0, a1 = 0, a0 = 0;
}
if (unlikely(c & 64)) {
sticky |= a3;
a3 = a2, a2 = a1, a1 = a0, a0 = 0;
}
c &= 63;
if (c == 0) {
goto done;
}
} else {
sticky = a0 | a1 | a2 | a3;
a0 = a1 = a2 = a3 = 0;
goto done;
}
sticky |= shr_double(a3, 0, c);
a3 = shr_double(a2, a3, c);
a2 = shr_double(a1, a2, c);
a1 = shr_double(a0, a1, c);
a0 = a0 >> c;
done:
a->frac_lo = a3 | (sticky != 0);
a->frac_lm = a2;
a->frac_hm = a1;
a->frac_hi = a0;
}
#define frac_shrjam(A, C) FRAC_GENERIC_64_128_256(shrjam, A)(A, C)
static bool frac64_sub(FloatParts64 *r, FloatParts64 *a, FloatParts64 *b)
{
return usub64_overflow(a->frac, b->frac, &r->frac);
}
static bool frac128_sub(FloatParts128 *r, FloatParts128 *a, FloatParts128 *b)
{
bool c = 0;
r->frac_lo = usub64_borrow(a->frac_lo, b->frac_lo, &c);
r->frac_hi = usub64_borrow(a->frac_hi, b->frac_hi, &c);
return c;
}
static bool frac256_sub(FloatParts256 *r, FloatParts256 *a, FloatParts256 *b)
{
bool c = 0;
r->frac_lo = usub64_borrow(a->frac_lo, b->frac_lo, &c);
r->frac_lm = usub64_borrow(a->frac_lm, b->frac_lm, &c);
r->frac_hm = usub64_borrow(a->frac_hm, b->frac_hm, &c);
r->frac_hi = usub64_borrow(a->frac_hi, b->frac_hi, &c);
return c;
}
#define frac_sub(R, A, B) FRAC_GENERIC_64_128_256(sub, R)(R, A, B)
static void frac64_truncjam(FloatParts64 *r, FloatParts128 *a)
{
r->frac = a->frac_hi | (a->frac_lo != 0);
}
static void frac128_truncjam(FloatParts128 *r, FloatParts256 *a)
{
r->frac_hi = a->frac_hi;
r->frac_lo = a->frac_hm | ((a->frac_lm | a->frac_lo) != 0);
}
#define frac_truncjam(R, A) FRAC_GENERIC_64_128(truncjam, R)(R, A)
static void frac64_widen(FloatParts128 *r, FloatParts64 *a)
{
r->frac_hi = a->frac;
r->frac_lo = 0;
}
static void frac128_widen(FloatParts256 *r, FloatParts128 *a)
{
r->frac_hi = a->frac_hi;
r->frac_hm = a->frac_lo;
r->frac_lm = 0;
r->frac_lo = 0;
}
#define frac_widen(A, B) FRAC_GENERIC_64_128(widen, B)(A, B)
/*
* Reciprocal sqrt table. 1 bit of exponent, 6-bits of mantessa.
* From https://git.musl-libc.org/cgit/musl/tree/src/math/sqrt_data.c
* and thus MIT licenced.
*/
static const uint16_t rsqrt_tab[128] = {
0xb451, 0xb2f0, 0xb196, 0xb044, 0xaef9, 0xadb6, 0xac79, 0xab43,
0xaa14, 0xa8eb, 0xa7c8, 0xa6aa, 0xa592, 0xa480, 0xa373, 0xa26b,
0xa168, 0xa06a, 0x9f70, 0x9e7b, 0x9d8a, 0x9c9d, 0x9bb5, 0x9ad1,
0x99f0, 0x9913, 0x983a, 0x9765, 0x9693, 0x95c4, 0x94f8, 0x9430,
0x936b, 0x92a9, 0x91ea, 0x912e, 0x9075, 0x8fbe, 0x8f0a, 0x8e59,
0x8daa, 0x8cfe, 0x8c54, 0x8bac, 0x8b07, 0x8a64, 0x89c4, 0x8925,
0x8889, 0x87ee, 0x8756, 0x86c0, 0x862b, 0x8599, 0x8508, 0x8479,
0x83ec, 0x8361, 0x82d8, 0x8250, 0x81c9, 0x8145, 0x80c2, 0x8040,
0xff02, 0xfd0e, 0xfb25, 0xf947, 0xf773, 0xf5aa, 0xf3ea, 0xf234,
0xf087, 0xeee3, 0xed47, 0xebb3, 0xea27, 0xe8a3, 0xe727, 0xe5b2,
0xe443, 0xe2dc, 0xe17a, 0xe020, 0xdecb, 0xdd7d, 0xdc34, 0xdaf1,
0xd9b3, 0xd87b, 0xd748, 0xd61a, 0xd4f1, 0xd3cd, 0xd2ad, 0xd192,
0xd07b, 0xcf69, 0xce5b, 0xcd51, 0xcc4a, 0xcb48, 0xca4a, 0xc94f,
0xc858, 0xc764, 0xc674, 0xc587, 0xc49d, 0xc3b7, 0xc2d4, 0xc1f4,
0xc116, 0xc03c, 0xbf65, 0xbe90, 0xbdbe, 0xbcef, 0xbc23, 0xbb59,
0xba91, 0xb9cc, 0xb90a, 0xb84a, 0xb78c, 0xb6d0, 0xb617, 0xb560,
};
#define partsN(NAME) glue(glue(glue(parts,N),_),NAME)
#define FloatPartsN glue(FloatParts,N)
#define FloatPartsW glue(FloatParts,W)
#define N 64
#define W 128
#include "softfloat-parts-addsub.c.inc"
#include "softfloat-parts.c.inc"
#undef N
#undef W
#define N 128
#define W 256
#include "softfloat-parts-addsub.c.inc"
#include "softfloat-parts.c.inc"
#undef N
#undef W
#define N 256
#include "softfloat-parts-addsub.c.inc"
#undef N
#undef W
#undef partsN
#undef FloatPartsN
#undef FloatPartsW
/*
* Pack/unpack routines with a specific FloatFmt.
*/
static void float16a_unpack_canonical(FloatParts64 *p, float16 f,
float_status *s, const FloatFmt *params)
{
float16_unpack_raw(p, f);
parts_canonicalize(p, s, params);
}
static void float16_unpack_canonical(FloatParts64 *p, float16 f,
float_status *s)
{
float16a_unpack_canonical(p, f, s, &float16_params);
}
static void bfloat16_unpack_canonical(FloatParts64 *p, bfloat16 f,
float_status *s)
{
bfloat16_unpack_raw(p, f);
parts_canonicalize(p, s, &bfloat16_params);
}
static float16 float16a_round_pack_canonical(FloatParts64 *p,
float_status *s,
const FloatFmt *params)
{
parts_uncanon(p, s, params);
return float16_pack_raw(p);
}
static float16 float16_round_pack_canonical(FloatParts64 *p,
float_status *s)
{
return float16a_round_pack_canonical(p, s, &float16_params);
}
static bfloat16 bfloat16_round_pack_canonical(FloatParts64 *p,
float_status *s)
{
parts_uncanon(p, s, &bfloat16_params);
return bfloat16_pack_raw(p);
}
static void float32_unpack_canonical(FloatParts64 *p, float32 f,
float_status *s)
{
float32_unpack_raw(p, f);
parts_canonicalize(p, s, &float32_params);
}
static float32 float32_round_pack_canonical(FloatParts64 *p,
float_status *s)
{
parts_uncanon(p, s, &float32_params);
return float32_pack_raw(p);
}
static void float64_unpack_canonical(FloatParts64 *p, float64 f,
float_status *s)
{
float64_unpack_raw(p, f);
parts_canonicalize(p, s, &float64_params);
}
static float64 float64_round_pack_canonical(FloatParts64 *p,
float_status *s)
{
parts_uncanon(p, s, &float64_params);
return float64_pack_raw(p);
}
static float64 float64r32_round_pack_canonical(FloatParts64 *p,
float_status *s)
{
parts_uncanon(p, s, &float32_params);
/*
* In parts_uncanon, we placed the fraction for float32 at the lsb.
* We need to adjust the fraction higher so that the least N bits are
* zero, and the fraction is adjacent to the float64 implicit bit.
*/
switch (p->cls) {
case float_class_normal:
if (unlikely(p->exp == 0)) {
/*
* The result is denormal for float32, but can be represented
* in normalized form for float64. Adjust, per canonicalize.
*/
int shift = frac_normalize(p);
p->exp = (float32_params.frac_shift -
float32_params.exp_bias - shift + 1 +
float64_params.exp_bias);
frac_shr(p, float64_params.frac_shift);
} else {
frac_shl(p, float32_params.frac_shift - float64_params.frac_shift);
p->exp += float64_params.exp_bias - float32_params.exp_bias;
}
break;
case float_class_snan:
case float_class_qnan:
frac_shl(p, float32_params.frac_shift - float64_params.frac_shift);
p->exp = float64_params.exp_max;
break;
case float_class_inf:
p->exp = float64_params.exp_max;
break;
case float_class_zero:
break;
default:
g_assert_not_reached();
}
return float64_pack_raw(p);
}
static void float128_unpack_canonical(FloatParts128 *p, float128 f,
float_status *s)
{
float128_unpack_raw(p, f);
parts_canonicalize(p, s, &float128_params);
}
static float128 float128_round_pack_canonical(FloatParts128 *p,
float_status *s)
{
parts_uncanon(p, s, &float128_params);
return float128_pack_raw(p);
}
/* Returns false if the encoding is invalid. */
static bool floatx80_unpack_canonical(FloatParts128 *p, floatx80 f,
float_status *s)
{
/* Ensure rounding precision is set before beginning. */
switch (s->floatx80_rounding_precision) {
case floatx80_precision_x:
case floatx80_precision_d:
case floatx80_precision_s:
break;
default:
g_assert_not_reached();
}
if (unlikely(floatx80_invalid_encoding(f))) {
float_raise(float_flag_invalid, s);
return false;
}
floatx80_unpack_raw(p, f);
if (likely(p->exp != floatx80_params[floatx80_precision_x].exp_max)) {
parts_canonicalize(p, s, &floatx80_params[floatx80_precision_x]);
} else {
/* The explicit integer bit is ignored, after invalid checks. */
p->frac_hi &= MAKE_64BIT_MASK(0, 63);
p->cls = (p->frac_hi == 0 ? float_class_inf
: parts_is_snan_frac(p->frac_hi, s)
? float_class_snan : float_class_qnan);
}
return true;
}
static floatx80 floatx80_round_pack_canonical(FloatParts128 *p,
float_status *s)
{
const FloatFmt *fmt = &floatx80_params[s->floatx80_rounding_precision];
uint64_t frac;
int exp;
switch (p->cls) {
case float_class_normal:
if (s->floatx80_rounding_precision == floatx80_precision_x) {
parts_uncanon_normal(p, s, fmt);
frac = p->frac_hi;
exp = p->exp;
} else {
FloatParts64 p64;
p64.sign = p->sign;
p64.exp = p->exp;
frac_truncjam(&p64, p);
parts_uncanon_normal(&p64, s, fmt);
frac = p64.frac;
exp = p64.exp;
}
if (exp != fmt->exp_max) {
break;
}
/* rounded to inf -- fall through to set frac correctly */
case float_class_inf:
/* x86 and m68k differ in the setting of the integer bit. */
frac = floatx80_infinity_low;
exp = fmt->exp_max;
break;
case float_class_zero:
frac = 0;
exp = 0;
break;
case float_class_snan:
case float_class_qnan:
/* NaNs have the integer bit set. */
frac = p->frac_hi | (1ull << 63);
exp = fmt->exp_max;
break;
default:
g_assert_not_reached();
}
return packFloatx80(p->sign, exp, frac);
}
/*
* Addition and subtraction
*/
static float16 QEMU_FLATTEN
float16_addsub(float16 a, float16 b, float_status *status, bool subtract)
{
FloatParts64 pa, pb, *pr;
float16_unpack_canonical(&pa, a, status);
float16_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return float16_round_pack_canonical(pr, status);
}
float16 float16_add(float16 a, float16 b, float_status *status)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
return float16_addsub(a, b, status, false);
}
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
float16 float16_sub(float16 a, float16 b, float_status *status)
{
return float16_addsub(a, b, status, true);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
static float32 QEMU_SOFTFLOAT_ATTR
soft_f32_addsub(float32 a, float32 b, float_status *status, bool subtract)
{
FloatParts64 pa, pb, *pr;
float32_unpack_canonical(&pa, a, status);
float32_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return float32_round_pack_canonical(pr, status);
}
static float32 soft_f32_add(float32 a, float32 b, float_status *status)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
return soft_f32_addsub(a, b, status, false);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
static float32 soft_f32_sub(float32 a, float32 b, float_status *status)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
return soft_f32_addsub(a, b, status, true);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_f64_addsub(float64 a, float64 b, float_status *status, bool subtract)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return float64_round_pack_canonical(pr, status);
}
static float64 soft_f64_add(float64 a, float64 b, float_status *status)
{
return soft_f64_addsub(a, b, status, false);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
static float64 soft_f64_sub(float64 a, float64 b, float_status *status)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
return soft_f64_addsub(a, b, status, true);
}
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
static float hard_f32_add(float a, float b)
{
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
return a + b;
}
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
static float hard_f32_sub(float a, float b)
{
return a - b;
}
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
static double hard_f64_add(double a, double b)
{
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
return a + b;
}
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
static double hard_f64_sub(double a, double b)
{
return a - b;
}
static bool f32_addsubmul_post(union_float32 a, union_float32 b)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
if (QEMU_HARDFLOAT_2F32_USE_FP) {
return !(fpclassify(a.h) == FP_ZERO && fpclassify(b.h) == FP_ZERO);
}
return !(float32_is_zero(a.s) && float32_is_zero(b.s));
}
static bool f64_addsubmul_post(union_float64 a, union_float64 b)
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
{
if (QEMU_HARDFLOAT_2F64_USE_FP) {
return !(fpclassify(a.h) == FP_ZERO && fpclassify(b.h) == FP_ZERO);
} else {
return !(float64_is_zero(a.s) && float64_is_zero(b.s));
}
}
static float32 float32_addsub(float32 a, float32 b, float_status *s,
hard_f32_op2_fn hard, soft_f32_op2_fn soft)
{
return float32_gen2(a, b, s, hard, soft,
f32_is_zon2, f32_addsubmul_post);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
static float64 float64_addsub(float64 a, float64 b, float_status *s,
hard_f64_op2_fn hard, soft_f64_op2_fn soft)
{
return float64_gen2(a, b, s, hard, soft,
f64_is_zon2, f64_addsubmul_post);
hardfloat: implement float32/64 addition and subtraction Performance results (single and double precision) for fp-bench: 1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: add-single: 135.07 MFlops add-double: 131.60 MFlops sub-single: 130.04 MFlops sub-double: 133.01 MFlops - after: add-single: 443.04 MFlops add-double: 301.95 MFlops sub-single: 411.36 MFlops sub-double: 293.15 MFlops 2. ARM Aarch64 A57 @ 2.4GHz - before: add-single: 44.79 MFlops add-double: 49.20 MFlops sub-single: 44.55 MFlops sub-double: 49.06 MFlops - after: add-single: 93.28 MFlops add-double: 88.27 MFlops sub-single: 91.47 MFlops sub-double: 88.27 MFlops 3. IBM POWER8E @ 2.1 GHz - before: add-single: 72.59 MFlops add-double: 72.27 MFlops sub-single: 75.33 MFlops sub-double: 70.54 MFlops - after: add-single: 112.95 MFlops add-double: 201.11 MFlops sub-single: 116.80 MFlops sub-double: 188.72 MFlops Note that the IBM and ARM machines benefit from having HARDFLOAT_2F{32,64}_USE_FP set to 0. Otherwise their performance can suffer significantly: - IBM Power8: add-single: [1] 54.94 vs [0] 116.37 MFlops add-double: [1] 58.92 vs [0] 201.44 MFlops - Aarch64 A57: add-single: [1] 80.72 vs [0] 93.24 MFlops add-double: [1] 82.10 vs [0] 88.18 MFlops On the Intel machine, having 2F64 set to 1 pays off, but it doesn't for 2F32: - Intel i7-6700K: add-single: [1] 285.79 vs [0] 426.70 MFlops add-double: [1] 302.15 vs [0] 278.82 MFlops Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 02:58:42 +03:00
}
float32 QEMU_FLATTEN
float32_add(float32 a, float32 b, float_status *s)
{
return float32_addsub(a, b, s, hard_f32_add, soft_f32_add);
}
float32 QEMU_FLATTEN
float32_sub(float32 a, float32 b, float_status *s)
{
return float32_addsub(a, b, s, hard_f32_sub, soft_f32_sub);
}
float64 QEMU_FLATTEN
float64_add(float64 a, float64 b, float_status *s)
{
return float64_addsub(a, b, s, hard_f64_add, soft_f64_add);
}
float64 QEMU_FLATTEN
float64_sub(float64 a, float64 b, float_status *s)
{
return float64_addsub(a, b, s, hard_f64_sub, soft_f64_sub);
}
static float64 float64r32_addsub(float64 a, float64 b, float_status *status,
bool subtract)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return float64r32_round_pack_canonical(pr, status);
}
float64 float64r32_add(float64 a, float64 b, float_status *status)
{
return float64r32_addsub(a, b, status, false);
}
float64 float64r32_sub(float64 a, float64 b, float_status *status)
{
return float64r32_addsub(a, b, status, true);
}
static bfloat16 QEMU_FLATTEN
bfloat16_addsub(bfloat16 a, bfloat16 b, float_status *status, bool subtract)
{
FloatParts64 pa, pb, *pr;
bfloat16_unpack_canonical(&pa, a, status);
bfloat16_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return bfloat16_round_pack_canonical(pr, status);
}
bfloat16 bfloat16_add(bfloat16 a, bfloat16 b, float_status *status)
{
return bfloat16_addsub(a, b, status, false);
}
bfloat16 bfloat16_sub(bfloat16 a, bfloat16 b, float_status *status)
{
return bfloat16_addsub(a, b, status, true);
}
static float128 QEMU_FLATTEN
float128_addsub(float128 a, float128 b, float_status *status, bool subtract)
{
FloatParts128 pa, pb, *pr;
float128_unpack_canonical(&pa, a, status);
float128_unpack_canonical(&pb, b, status);
pr = parts_addsub(&pa, &pb, status, subtract);
return float128_round_pack_canonical(pr, status);
}
float128 float128_add(float128 a, float128 b, float_status *status)
{
return float128_addsub(a, b, status, false);
}
float128 float128_sub(float128 a, float128 b, float_status *status)
{
return float128_addsub(a, b, status, true);
}
static floatx80 QEMU_FLATTEN
floatx80_addsub(floatx80 a, floatx80 b, float_status *status, bool subtract)
{
FloatParts128 pa, pb, *pr;
if (!floatx80_unpack_canonical(&pa, a, status) ||
!floatx80_unpack_canonical(&pb, b, status)) {
return floatx80_default_nan(status);
}
pr = parts_addsub(&pa, &pb, status, subtract);
return floatx80_round_pack_canonical(pr, status);
}
floatx80 floatx80_add(floatx80 a, floatx80 b, float_status *status)
{
return floatx80_addsub(a, b, status, false);
}
floatx80 floatx80_sub(floatx80 a, floatx80 b, float_status *status)
{
return floatx80_addsub(a, b, status, true);
}
/*
* Multiplication
*/
float16 QEMU_FLATTEN float16_mul(float16 a, float16 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float16_unpack_canonical(&pa, a, status);
float16_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return float16_round_pack_canonical(pr, status);
}
static float32 QEMU_SOFTFLOAT_ATTR
soft_f32_mul(float32 a, float32 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float32_unpack_canonical(&pa, a, status);
float32_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return float32_round_pack_canonical(pr, status);
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_f64_mul(float64 a, float64 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return float64_round_pack_canonical(pr, status);
}
static float hard_f32_mul(float a, float b)
{
return a * b;
}
static double hard_f64_mul(double a, double b)
{
return a * b;
}
float32 QEMU_FLATTEN
float32_mul(float32 a, float32 b, float_status *s)
{
return float32_gen2(a, b, s, hard_f32_mul, soft_f32_mul,
f32_is_zon2, f32_addsubmul_post);
}
float64 QEMU_FLATTEN
float64_mul(float64 a, float64 b, float_status *s)
{
return float64_gen2(a, b, s, hard_f64_mul, soft_f64_mul,
f64_is_zon2, f64_addsubmul_post);
}
float64 float64r32_mul(float64 a, float64 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return float64r32_round_pack_canonical(pr, status);
}
bfloat16 QEMU_FLATTEN
bfloat16_mul(bfloat16 a, bfloat16 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
bfloat16_unpack_canonical(&pa, a, status);
bfloat16_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return bfloat16_round_pack_canonical(pr, status);
}
float128 QEMU_FLATTEN
float128_mul(float128 a, float128 b, float_status *status)
{
FloatParts128 pa, pb, *pr;
float128_unpack_canonical(&pa, a, status);
float128_unpack_canonical(&pb, b, status);
pr = parts_mul(&pa, &pb, status);
return float128_round_pack_canonical(pr, status);
}
floatx80 QEMU_FLATTEN
floatx80_mul(floatx80 a, floatx80 b, float_status *status)
{
FloatParts128 pa, pb, *pr;
if (!floatx80_unpack_canonical(&pa, a, status) ||
!floatx80_unpack_canonical(&pb, b, status)) {
return floatx80_default_nan(status);
}
pr = parts_mul(&pa, &pb, status);
return floatx80_round_pack_canonical(pr, status);
}
/*
* Fused multiply-add
*/
float16 QEMU_FLATTEN float16_muladd(float16 a, float16 b, float16 c,
int flags, float_status *status)
{
FloatParts64 pa, pb, pc, *pr;
float16_unpack_canonical(&pa, a, status);
float16_unpack_canonical(&pb, b, status);
float16_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return float16_round_pack_canonical(pr, status);
}
static float32 QEMU_SOFTFLOAT_ATTR
soft_f32_muladd(float32 a, float32 b, float32 c, int flags,
float_status *status)
{
FloatParts64 pa, pb, pc, *pr;
float32_unpack_canonical(&pa, a, status);
float32_unpack_canonical(&pb, b, status);
float32_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return float32_round_pack_canonical(pr, status);
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_f64_muladd(float64 a, float64 b, float64 c, int flags,
float_status *status)
{
FloatParts64 pa, pb, pc, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
float64_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return float64_round_pack_canonical(pr, status);
}
static bool force_soft_fma;
float32 QEMU_FLATTEN
float32_muladd(float32 xa, float32 xb, float32 xc, int flags, float_status *s)
{
union_float32 ua, ub, uc, ur;
ua.s = xa;
ub.s = xb;
uc.s = xc;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
if (unlikely(flags & float_muladd_halve_result)) {
goto soft;
}
float32_input_flush3(&ua.s, &ub.s, &uc.s, s);
if (unlikely(!f32_is_zon3(ua, ub, uc))) {
goto soft;
}
if (unlikely(force_soft_fma)) {
goto soft;
}
/*
* When (a || b) == 0, there's no need to check for under/over flow,
* since we know the addend is (normal || 0) and the product is 0.
*/
if (float32_is_zero(ua.s) || float32_is_zero(ub.s)) {
union_float32 up;
bool prod_sign;
prod_sign = float32_is_neg(ua.s) ^ float32_is_neg(ub.s);
prod_sign ^= !!(flags & float_muladd_negate_product);
up.s = float32_set_sign(float32_zero, prod_sign);
if (flags & float_muladd_negate_c) {
uc.h = -uc.h;
}
ur.h = up.h + uc.h;
} else {
union_float32 ua_orig = ua;
union_float32 uc_orig = uc;
if (flags & float_muladd_negate_product) {
ua.h = -ua.h;
}
if (flags & float_muladd_negate_c) {
uc.h = -uc.h;
}
ur.h = fmaf(ua.h, ub.h, uc.h);
if (unlikely(f32_is_inf(ur))) {
float_raise(float_flag_overflow, s);
} else if (unlikely(fabsf(ur.h) <= FLT_MIN)) {
ua = ua_orig;
uc = uc_orig;
goto soft;
}
}
if (flags & float_muladd_negate_result) {
return float32_chs(ur.s);
}
return ur.s;
soft:
return soft_f32_muladd(ua.s, ub.s, uc.s, flags, s);
}
float64 QEMU_FLATTEN
float64_muladd(float64 xa, float64 xb, float64 xc, int flags, float_status *s)
{
union_float64 ua, ub, uc, ur;
ua.s = xa;
ub.s = xb;
uc.s = xc;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
if (unlikely(flags & float_muladd_halve_result)) {
goto soft;
}
float64_input_flush3(&ua.s, &ub.s, &uc.s, s);
if (unlikely(!f64_is_zon3(ua, ub, uc))) {
goto soft;
}
if (unlikely(force_soft_fma)) {
goto soft;
}
/*
* When (a || b) == 0, there's no need to check for under/over flow,
* since we know the addend is (normal || 0) and the product is 0.
*/
if (float64_is_zero(ua.s) || float64_is_zero(ub.s)) {
union_float64 up;
bool prod_sign;
prod_sign = float64_is_neg(ua.s) ^ float64_is_neg(ub.s);
prod_sign ^= !!(flags & float_muladd_negate_product);
up.s = float64_set_sign(float64_zero, prod_sign);
if (flags & float_muladd_negate_c) {
uc.h = -uc.h;
}
ur.h = up.h + uc.h;
} else {
union_float64 ua_orig = ua;
union_float64 uc_orig = uc;
if (flags & float_muladd_negate_product) {
ua.h = -ua.h;
}
if (flags & float_muladd_negate_c) {
uc.h = -uc.h;
}
ur.h = fma(ua.h, ub.h, uc.h);
if (unlikely(f64_is_inf(ur))) {
float_raise(float_flag_overflow, s);
} else if (unlikely(fabs(ur.h) <= FLT_MIN)) {
ua = ua_orig;
uc = uc_orig;
goto soft;
}
}
if (flags & float_muladd_negate_result) {
return float64_chs(ur.s);
}
return ur.s;
soft:
return soft_f64_muladd(ua.s, ub.s, uc.s, flags, s);
}
float64 float64r32_muladd(float64 a, float64 b, float64 c,
int flags, float_status *status)
{
FloatParts64 pa, pb, pc, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
float64_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return float64r32_round_pack_canonical(pr, status);
}
bfloat16 QEMU_FLATTEN bfloat16_muladd(bfloat16 a, bfloat16 b, bfloat16 c,
int flags, float_status *status)
{
FloatParts64 pa, pb, pc, *pr;
bfloat16_unpack_canonical(&pa, a, status);
bfloat16_unpack_canonical(&pb, b, status);
bfloat16_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return bfloat16_round_pack_canonical(pr, status);
}
float128 QEMU_FLATTEN float128_muladd(float128 a, float128 b, float128 c,
int flags, float_status *status)
{
FloatParts128 pa, pb, pc, *pr;
float128_unpack_canonical(&pa, a, status);
float128_unpack_canonical(&pb, b, status);
float128_unpack_canonical(&pc, c, status);
pr = parts_muladd(&pa, &pb, &pc, flags, status);
return float128_round_pack_canonical(pr, status);
}
/*
* Division
*/
float16 float16_div(float16 a, float16 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float16_unpack_canonical(&pa, a, status);
float16_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return float16_round_pack_canonical(pr, status);
}
static float32 QEMU_SOFTFLOAT_ATTR
soft_f32_div(float32 a, float32 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float32_unpack_canonical(&pa, a, status);
float32_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return float32_round_pack_canonical(pr, status);
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_f64_div(float64 a, float64 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return float64_round_pack_canonical(pr, status);
}
static float hard_f32_div(float a, float b)
{
return a / b;
}
static double hard_f64_div(double a, double b)
{
return a / b;
}
static bool f32_div_pre(union_float32 a, union_float32 b)
{
if (QEMU_HARDFLOAT_2F32_USE_FP) {
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
fpclassify(b.h) == FP_NORMAL;
}
return float32_is_zero_or_normal(a.s) && float32_is_normal(b.s);
}
static bool f64_div_pre(union_float64 a, union_float64 b)
{
if (QEMU_HARDFLOAT_2F64_USE_FP) {
return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) &&
fpclassify(b.h) == FP_NORMAL;
}
return float64_is_zero_or_normal(a.s) && float64_is_normal(b.s);
}
static bool f32_div_post(union_float32 a, union_float32 b)
{
if (QEMU_HARDFLOAT_2F32_USE_FP) {
return fpclassify(a.h) != FP_ZERO;
}
return !float32_is_zero(a.s);
}
static bool f64_div_post(union_float64 a, union_float64 b)
{
if (QEMU_HARDFLOAT_2F64_USE_FP) {
return fpclassify(a.h) != FP_ZERO;
}
return !float64_is_zero(a.s);
}
float32 QEMU_FLATTEN
float32_div(float32 a, float32 b, float_status *s)
{
return float32_gen2(a, b, s, hard_f32_div, soft_f32_div,
f32_div_pre, f32_div_post);
}
float64 QEMU_FLATTEN
float64_div(float64 a, float64 b, float_status *s)
{
return float64_gen2(a, b, s, hard_f64_div, soft_f64_div,
f64_div_pre, f64_div_post);
}
float64 float64r32_div(float64 a, float64 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return float64r32_round_pack_canonical(pr, status);
}
bfloat16 QEMU_FLATTEN
bfloat16_div(bfloat16 a, bfloat16 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
bfloat16_unpack_canonical(&pa, a, status);
bfloat16_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return bfloat16_round_pack_canonical(pr, status);
}
float128 QEMU_FLATTEN
float128_div(float128 a, float128 b, float_status *status)
{
FloatParts128 pa, pb, *pr;
float128_unpack_canonical(&pa, a, status);
float128_unpack_canonical(&pb, b, status);
pr = parts_div(&pa, &pb, status);
return float128_round_pack_canonical(pr, status);
}
floatx80 floatx80_div(floatx80 a, floatx80 b, float_status *status)
{
FloatParts128 pa, pb, *pr;
if (!floatx80_unpack_canonical(&pa, a, status) ||
!floatx80_unpack_canonical(&pb, b, status)) {
return floatx80_default_nan(status);
}
pr = parts_div(&pa, &pb, status);
return floatx80_round_pack_canonical(pr, status);
}
/*
* Remainder
*/
float32 float32_rem(float32 a, float32 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float32_unpack_canonical(&pa, a, status);
float32_unpack_canonical(&pb, b, status);
pr = parts_modrem(&pa, &pb, NULL, status);
return float32_round_pack_canonical(pr, status);
}
float64 float64_rem(float64 a, float64 b, float_status *status)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, status);
float64_unpack_canonical(&pb, b, status);
pr = parts_modrem(&pa, &pb, NULL, status);
return float64_round_pack_canonical(pr, status);
}
float128 float128_rem(float128 a, float128 b, float_status *status)
{
FloatParts128 pa, pb, *pr;
float128_unpack_canonical(&pa, a, status);
float128_unpack_canonical(&pb, b, status);
pr = parts_modrem(&pa, &pb, NULL, status);
return float128_round_pack_canonical(pr, status);
}
/*
* Returns the remainder of the extended double-precision floating-point value
* `a' with respect to the corresponding value `b'.
* If 'mod' is false, the operation is performed according to the IEC/IEEE
* Standard for Binary Floating-Point Arithmetic. If 'mod' is true, return
* the remainder based on truncating the quotient toward zero instead and
* *quotient is set to the low 64 bits of the absolute value of the integer
* quotient.
*/
floatx80 floatx80_modrem(floatx80 a, floatx80 b, bool mod,
uint64_t *quotient, float_status *status)
{
FloatParts128 pa, pb, *pr;
*quotient = 0;
if (!floatx80_unpack_canonical(&pa, a, status) ||
!floatx80_unpack_canonical(&pb, b, status)) {
return floatx80_default_nan(status);
}
pr = parts_modrem(&pa, &pb, mod ? quotient : NULL, status);
return floatx80_round_pack_canonical(pr, status);
}
floatx80 floatx80_rem(floatx80 a, floatx80 b, float_status *status)
{
uint64_t quotient;
return floatx80_modrem(a, b, false, &quotient, status);
}
floatx80 floatx80_mod(floatx80 a, floatx80 b, float_status *status)
{
uint64_t quotient;
return floatx80_modrem(a, b, true, &quotient, status);
}
/*
* Float to Float conversions
*
* Returns the result of converting one float format to another. The
* conversion is performed according to the IEC/IEEE Standard for
* Binary Floating-Point Arithmetic.
*
* Usually this only needs to take care of raising invalid exceptions
* and handling the conversion on NaNs.
*/
static void parts_float_to_ahp(FloatParts64 *a, float_status *s)
{
switch (a->cls) {
case float_class_snan:
float_raise(float_flag_invalid_snan, s);
/* fall through */
case float_class_qnan:
/*
* There is no NaN in the destination format. Raise Invalid
* and return a zero with the sign of the input NaN.
*/
float_raise(float_flag_invalid, s);
a->cls = float_class_zero;
break;
case float_class_inf:
/*
* There is no Inf in the destination format. Raise Invalid
* and return the maximum normal with the correct sign.
*/
float_raise(float_flag_invalid, s);
a->cls = float_class_normal;
a->exp = float16_params_ahp.exp_max;
a->frac = MAKE_64BIT_MASK(float16_params_ahp.frac_shift,
float16_params_ahp.frac_size + 1);
break;
case float_class_normal:
case float_class_zero:
break;
default:
g_assert_not_reached();
}
}
static void parts64_float_to_float(FloatParts64 *a, float_status *s)
{
if (is_nan(a->cls)) {
parts_return_nan(a, s);
}
}
static void parts128_float_to_float(FloatParts128 *a, float_status *s)
{
if (is_nan(a->cls)) {
parts_return_nan(a, s);
}
}
#define parts_float_to_float(P, S) \
PARTS_GENERIC_64_128(float_to_float, P)(P, S)
static void parts_float_to_float_narrow(FloatParts64 *a, FloatParts128 *b,
float_status *s)
{
a->cls = b->cls;
a->sign = b->sign;
a->exp = b->exp;
if (a->cls == float_class_normal) {
frac_truncjam(a, b);
} else if (is_nan(a->cls)) {
/* Discard the low bits of the NaN. */
a->frac = b->frac_hi;
parts_return_nan(a, s);
}
}
static void parts_float_to_float_widen(FloatParts128 *a, FloatParts64 *b,
float_status *s)
{
a->cls = b->cls;
a->sign = b->sign;
a->exp = b->exp;
frac_widen(a, b);
if (is_nan(a->cls)) {
parts_return_nan(a, s);
}
}
float32 float16_to_float32(float16 a, bool ieee, float_status *s)
{
const FloatFmt *fmt16 = ieee ? &float16_params : &float16_params_ahp;
FloatParts64 p;
float16a_unpack_canonical(&p, a, s, fmt16);
parts_float_to_float(&p, s);
return float32_round_pack_canonical(&p, s);
}
float64 float16_to_float64(float16 a, bool ieee, float_status *s)
{
const FloatFmt *fmt16 = ieee ? &float16_params : &float16_params_ahp;
FloatParts64 p;
float16a_unpack_canonical(&p, a, s, fmt16);
parts_float_to_float(&p, s);
return float64_round_pack_canonical(&p, s);
}
float16 float32_to_float16(float32 a, bool ieee, float_status *s)
{
FloatParts64 p;
const FloatFmt *fmt;
float32_unpack_canonical(&p, a, s);
if (ieee) {
parts_float_to_float(&p, s);
fmt = &float16_params;
} else {
parts_float_to_ahp(&p, s);
fmt = &float16_params_ahp;
}
return float16a_round_pack_canonical(&p, s, fmt);
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_float32_to_float64(float32 a, float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return float64_round_pack_canonical(&p, s);
}
float64 float32_to_float64(float32 a, float_status *s)
{
if (likely(float32_is_normal(a))) {
/* Widening conversion can never produce inexact results. */
union_float32 uf;
union_float64 ud;
uf.s = a;
ud.h = uf.h;
return ud.s;
} else if (float32_is_zero(a)) {
return float64_set_sign(float64_zero, float32_is_neg(a));
} else {
return soft_float32_to_float64(a, s);
}
}
float16 float64_to_float16(float64 a, bool ieee, float_status *s)
{
FloatParts64 p;
const FloatFmt *fmt;
float64_unpack_canonical(&p, a, s);
if (ieee) {
parts_float_to_float(&p, s);
fmt = &float16_params;
} else {
parts_float_to_ahp(&p, s);
fmt = &float16_params_ahp;
}
return float16a_round_pack_canonical(&p, s, fmt);
}
float32 float64_to_float32(float64 a, float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return float32_round_pack_canonical(&p, s);
}
float32 bfloat16_to_float32(bfloat16 a, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return float32_round_pack_canonical(&p, s);
}
float64 bfloat16_to_float64(bfloat16 a, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return float64_round_pack_canonical(&p, s);
}
bfloat16 float32_to_bfloat16(float32 a, float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return bfloat16_round_pack_canonical(&p, s);
}
bfloat16 float64_to_bfloat16(float64 a, float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return bfloat16_round_pack_canonical(&p, s);
}
float32 float128_to_float32(float128 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float128_unpack_canonical(&p128, a, s);
parts_float_to_float_narrow(&p64, &p128, s);
return float32_round_pack_canonical(&p64, s);
}
float64 float128_to_float64(float128 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float128_unpack_canonical(&p128, a, s);
parts_float_to_float_narrow(&p64, &p128, s);
return float64_round_pack_canonical(&p64, s);
}
float128 float32_to_float128(float32 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float32_unpack_canonical(&p64, a, s);
parts_float_to_float_widen(&p128, &p64, s);
return float128_round_pack_canonical(&p128, s);
}
float128 float64_to_float128(float64 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float64_unpack_canonical(&p64, a, s);
parts_float_to_float_widen(&p128, &p64, s);
return float128_round_pack_canonical(&p128, s);
}
float32 floatx80_to_float32(floatx80 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
if (floatx80_unpack_canonical(&p128, a, s)) {
parts_float_to_float_narrow(&p64, &p128, s);
} else {
parts_default_nan(&p64, s);
}
return float32_round_pack_canonical(&p64, s);
}
float64 floatx80_to_float64(floatx80 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
if (floatx80_unpack_canonical(&p128, a, s)) {
parts_float_to_float_narrow(&p64, &p128, s);
} else {
parts_default_nan(&p64, s);
}
return float64_round_pack_canonical(&p64, s);
}
float128 floatx80_to_float128(floatx80 a, float_status *s)
{
FloatParts128 p;
if (floatx80_unpack_canonical(&p, a, s)) {
parts_float_to_float(&p, s);
} else {
parts_default_nan(&p, s);
}
return float128_round_pack_canonical(&p, s);
}
floatx80 float32_to_floatx80(float32 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float32_unpack_canonical(&p64, a, s);
parts_float_to_float_widen(&p128, &p64, s);
return floatx80_round_pack_canonical(&p128, s);
}
floatx80 float64_to_floatx80(float64 a, float_status *s)
{
FloatParts64 p64;
FloatParts128 p128;
float64_unpack_canonical(&p64, a, s);
parts_float_to_float_widen(&p128, &p64, s);
return floatx80_round_pack_canonical(&p128, s);
}
floatx80 float128_to_floatx80(float128 a, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
parts_float_to_float(&p, s);
return floatx80_round_pack_canonical(&p, s);
}
/*
* Round to integral value
*/
float16 float16_round_to_int(float16 a, float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float16_params);
return float16_round_pack_canonical(&p, s);
}
float32 float32_round_to_int(float32 a, float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float32_params);
return float32_round_pack_canonical(&p, s);
}
float64 float64_round_to_int(float64 a, float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float64_params);
return float64_round_pack_canonical(&p, s);
}
bfloat16 bfloat16_round_to_int(bfloat16 a, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
parts_round_to_int(&p, s->float_rounding_mode, 0, s, &bfloat16_params);
return bfloat16_round_pack_canonical(&p, s);
}
float128 float128_round_to_int(float128 a, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float128_params);
return float128_round_pack_canonical(&p, s);
}
floatx80 floatx80_round_to_int(floatx80 a, float_status *status)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, status)) {
return floatx80_default_nan(status);
}
parts_round_to_int(&p, status->float_rounding_mode, 0, status,
&floatx80_params[status->floatx80_rounding_precision]);
return floatx80_round_pack_canonical(&p, status);
}
/*
* Floating-point to signed integer conversions
*/
int8_t float16_to_int8_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT8_MIN, INT8_MAX, s);
}
int16_t float16_to_int16_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s);
}
int32_t float16_to_int32_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
int64_t float16_to_int64_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
int16_t float32_to_int16_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s);
}
int32_t float32_to_int32_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
int64_t float32_to_int64_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
int16_t float64_to_int16_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s);
}
int32_t float64_to_int32_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
int64_t float64_to_int64_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
int16_t bfloat16_to_int16_scalbn(bfloat16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s);
}
int32_t bfloat16_to_int32_scalbn(bfloat16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
int64_t bfloat16_to_int64_scalbn(bfloat16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
static int32_t float128_to_int32_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
static int64_t float128_to_int64_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
static Int128 float128_to_int128_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
int flags = 0;
Int128 r;
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
switch (p.cls) {
case float_class_snan:
flags |= float_flag_invalid_snan;
/* fall through */
case float_class_qnan:
flags |= float_flag_invalid;
r = UINT128_MAX;
break;
case float_class_inf:
flags = float_flag_invalid | float_flag_invalid_cvti;
r = p.sign ? INT128_MIN : INT128_MAX;
break;
case float_class_zero:
return int128_zero();
case float_class_normal:
if (parts_round_to_int_normal(&p, rmode, scale, 128 - 2)) {
flags = float_flag_inexact;
}
if (p.exp < 127) {
int shift = 127 - p.exp;
r = int128_urshift(int128_make128(p.frac_lo, p.frac_hi), shift);
if (p.sign) {
r = int128_neg(r);
}
} else if (p.exp == 127 && p.sign && p.frac_lo == 0 &&
p.frac_hi == DECOMPOSED_IMPLICIT_BIT) {
r = INT128_MIN;
} else {
flags = float_flag_invalid | float_flag_invalid_cvti;
r = p.sign ? INT128_MIN : INT128_MAX;
}
break;
default:
g_assert_not_reached();
}
float_raise(flags, s);
return r;
}
static int32_t floatx80_to_int32_scalbn(floatx80 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, s)) {
parts_default_nan(&p, s);
}
return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s);
}
static int64_t floatx80_to_int64_scalbn(floatx80 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, s)) {
parts_default_nan(&p, s);
}
return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s);
}
int8_t float16_to_int8(float16 a, float_status *s)
{
return float16_to_int8_scalbn(a, s->float_rounding_mode, 0, s);
}
int16_t float16_to_int16(float16 a, float_status *s)
{
return float16_to_int16_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t float16_to_int32(float16 a, float_status *s)
{
return float16_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t float16_to_int64(float16 a, float_status *s)
{
return float16_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
int16_t float32_to_int16(float32 a, float_status *s)
{
return float32_to_int16_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t float32_to_int32(float32 a, float_status *s)
{
return float32_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t float32_to_int64(float32 a, float_status *s)
{
return float32_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
int16_t float64_to_int16(float64 a, float_status *s)
{
return float64_to_int16_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t float64_to_int32(float64 a, float_status *s)
{
return float64_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t float64_to_int64(float64 a, float_status *s)
{
return float64_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t float128_to_int32(float128 a, float_status *s)
{
return float128_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t float128_to_int64(float128 a, float_status *s)
{
return float128_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
Int128 float128_to_int128(float128 a, float_status *s)
{
return float128_to_int128_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t floatx80_to_int32(floatx80 a, float_status *s)
{
return floatx80_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t floatx80_to_int64(floatx80 a, float_status *s)
{
return floatx80_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
int16_t float16_to_int16_round_to_zero(float16 a, float_status *s)
{
return float16_to_int16_scalbn(a, float_round_to_zero, 0, s);
}
int32_t float16_to_int32_round_to_zero(float16 a, float_status *s)
{
return float16_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t float16_to_int64_round_to_zero(float16 a, float_status *s)
{
return float16_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
int16_t float32_to_int16_round_to_zero(float32 a, float_status *s)
{
return float32_to_int16_scalbn(a, float_round_to_zero, 0, s);
}
int32_t float32_to_int32_round_to_zero(float32 a, float_status *s)
{
return float32_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t float32_to_int64_round_to_zero(float32 a, float_status *s)
{
return float32_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
int16_t float64_to_int16_round_to_zero(float64 a, float_status *s)
{
return float64_to_int16_scalbn(a, float_round_to_zero, 0, s);
}
int32_t float64_to_int32_round_to_zero(float64 a, float_status *s)
{
return float64_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t float64_to_int64_round_to_zero(float64 a, float_status *s)
{
return float64_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
int32_t float128_to_int32_round_to_zero(float128 a, float_status *s)
{
return float128_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t float128_to_int64_round_to_zero(float128 a, float_status *s)
{
return float128_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
Int128 float128_to_int128_round_to_zero(float128 a, float_status *s)
{
return float128_to_int128_scalbn(a, float_round_to_zero, 0, s);
}
int32_t floatx80_to_int32_round_to_zero(floatx80 a, float_status *s)
{
return floatx80_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t floatx80_to_int64_round_to_zero(floatx80 a, float_status *s)
{
return floatx80_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
int16_t bfloat16_to_int16(bfloat16 a, float_status *s)
{
return bfloat16_to_int16_scalbn(a, s->float_rounding_mode, 0, s);
}
int32_t bfloat16_to_int32(bfloat16 a, float_status *s)
{
return bfloat16_to_int32_scalbn(a, s->float_rounding_mode, 0, s);
}
int64_t bfloat16_to_int64(bfloat16 a, float_status *s)
{
return bfloat16_to_int64_scalbn(a, s->float_rounding_mode, 0, s);
}
int16_t bfloat16_to_int16_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_int16_scalbn(a, float_round_to_zero, 0, s);
}
int32_t bfloat16_to_int32_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_int32_scalbn(a, float_round_to_zero, 0, s);
}
int64_t bfloat16_to_int64_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_int64_scalbn(a, float_round_to_zero, 0, s);
}
int32_t float64_to_int32_modulo(float64 a, FloatRoundMode rmode,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_sint_modulo(&p, rmode, 31, s);
}
int64_t float64_to_int64_modulo(float64 a, FloatRoundMode rmode,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_sint_modulo(&p, rmode, 63, s);
}
/*
* Floating-point to unsigned integer conversions
*/
uint8_t float16_to_uint8_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT8_MAX, s);
}
uint16_t float16_to_uint16_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s);
}
uint32_t float16_to_uint32_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s);
}
uint64_t float16_to_uint64_scalbn(float16 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s);
}
uint16_t float32_to_uint16_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s);
}
uint32_t float32_to_uint32_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s);
}
uint64_t float32_to_uint64_scalbn(float32 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s);
}
uint16_t float64_to_uint16_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s);
}
uint32_t float64_to_uint32_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s);
}
uint64_t float64_to_uint64_scalbn(float64 a, FloatRoundMode rmode, int scale,
float_status *s)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s);
}
uint16_t bfloat16_to_uint16_scalbn(bfloat16 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s);
}
uint32_t bfloat16_to_uint32_scalbn(bfloat16 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s);
}
uint64_t bfloat16_to_uint64_scalbn(bfloat16 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s);
}
static uint32_t float128_to_uint32_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s);
}
static uint64_t float128_to_uint64_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s);
}
static Int128 float128_to_uint128_scalbn(float128 a, FloatRoundMode rmode,
int scale, float_status *s)
{
int flags = 0;
Int128 r;
FloatParts128 p;
float128_unpack_canonical(&p, a, s);
switch (p.cls) {
case float_class_snan:
flags |= float_flag_invalid_snan;
/* fall through */
case float_class_qnan:
flags |= float_flag_invalid;
r = UINT128_MAX;
break;
case float_class_inf:
flags = float_flag_invalid | float_flag_invalid_cvti;
r = p.sign ? int128_zero() : UINT128_MAX;
break;
case float_class_zero:
return int128_zero();
case float_class_normal:
if (parts_round_to_int_normal(&p, rmode, scale, 128 - 2)) {
flags = float_flag_inexact;
if (p.cls == float_class_zero) {
r = int128_zero();
break;
}
}
if (p.sign) {
flags = float_flag_invalid | float_flag_invalid_cvti;
r = int128_zero();
} else if (p.exp <= 127) {
int shift = 127 - p.exp;
r = int128_urshift(int128_make128(p.frac_lo, p.frac_hi), shift);
} else {
flags = float_flag_invalid | float_flag_invalid_cvti;
r = UINT128_MAX;
}
break;
default:
g_assert_not_reached();
}
float_raise(flags, s);
return r;
}
uint8_t float16_to_uint8(float16 a, float_status *s)
{
return float16_to_uint8_scalbn(a, s->float_rounding_mode, 0, s);
}
uint16_t float16_to_uint16(float16 a, float_status *s)
{
return float16_to_uint16_scalbn(a, s->float_rounding_mode, 0, s);
}
uint32_t float16_to_uint32(float16 a, float_status *s)
{
return float16_to_uint32_scalbn(a, s->float_rounding_mode, 0, s);
}
uint64_t float16_to_uint64(float16 a, float_status *s)
{
return float16_to_uint64_scalbn(a, s->float_rounding_mode, 0, s);
}
uint16_t float32_to_uint16(float32 a, float_status *s)
{
return float32_to_uint16_scalbn(a, s->float_rounding_mode, 0, s);
}
uint32_t float32_to_uint32(float32 a, float_status *s)
{
return float32_to_uint32_scalbn(a, s->float_rounding_mode, 0, s);
}
uint64_t float32_to_uint64(float32 a, float_status *s)
{
return float32_to_uint64_scalbn(a, s->float_rounding_mode, 0, s);
}
uint16_t float64_to_uint16(float64 a, float_status *s)
{
return float64_to_uint16_scalbn(a, s->float_rounding_mode, 0, s);
}
uint32_t float64_to_uint32(float64 a, float_status *s)
{
return float64_to_uint32_scalbn(a, s->float_rounding_mode, 0, s);
}
uint64_t float64_to_uint64(float64 a, float_status *s)
{
return float64_to_uint64_scalbn(a, s->float_rounding_mode, 0, s);
}
uint32_t float128_to_uint32(float128 a, float_status *s)
{
return float128_to_uint32_scalbn(a, s->float_rounding_mode, 0, s);
}
uint64_t float128_to_uint64(float128 a, float_status *s)
{
return float128_to_uint64_scalbn(a, s->float_rounding_mode, 0, s);
}
Int128 float128_to_uint128(float128 a, float_status *s)
{
return float128_to_uint128_scalbn(a, s->float_rounding_mode, 0, s);
}
uint16_t float16_to_uint16_round_to_zero(float16 a, float_status *s)
{
return float16_to_uint16_scalbn(a, float_round_to_zero, 0, s);
}
uint32_t float16_to_uint32_round_to_zero(float16 a, float_status *s)
{
return float16_to_uint32_scalbn(a, float_round_to_zero, 0, s);
}
uint64_t float16_to_uint64_round_to_zero(float16 a, float_status *s)
{
return float16_to_uint64_scalbn(a, float_round_to_zero, 0, s);
}
uint16_t float32_to_uint16_round_to_zero(float32 a, float_status *s)
{
return float32_to_uint16_scalbn(a, float_round_to_zero, 0, s);
}
uint32_t float32_to_uint32_round_to_zero(float32 a, float_status *s)
{
return float32_to_uint32_scalbn(a, float_round_to_zero, 0, s);
}
uint64_t float32_to_uint64_round_to_zero(float32 a, float_status *s)
{
return float32_to_uint64_scalbn(a, float_round_to_zero, 0, s);
}
uint16_t float64_to_uint16_round_to_zero(float64 a, float_status *s)
{
return float64_to_uint16_scalbn(a, float_round_to_zero, 0, s);
}
uint32_t float64_to_uint32_round_to_zero(float64 a, float_status *s)
{
return float64_to_uint32_scalbn(a, float_round_to_zero, 0, s);
}
uint64_t float64_to_uint64_round_to_zero(float64 a, float_status *s)
{
return float64_to_uint64_scalbn(a, float_round_to_zero, 0, s);
}
uint32_t float128_to_uint32_round_to_zero(float128 a, float_status *s)
{
return float128_to_uint32_scalbn(a, float_round_to_zero, 0, s);
}
uint64_t float128_to_uint64_round_to_zero(float128 a, float_status *s)
{
return float128_to_uint64_scalbn(a, float_round_to_zero, 0, s);
}
Int128 float128_to_uint128_round_to_zero(float128 a, float_status *s)
{
return float128_to_uint128_scalbn(a, float_round_to_zero, 0, s);
}
uint16_t bfloat16_to_uint16(bfloat16 a, float_status *s)
{
return bfloat16_to_uint16_scalbn(a, s->float_rounding_mode, 0, s);
}
uint32_t bfloat16_to_uint32(bfloat16 a, float_status *s)
{
return bfloat16_to_uint32_scalbn(a, s->float_rounding_mode, 0, s);
}
uint64_t bfloat16_to_uint64(bfloat16 a, float_status *s)
{
return bfloat16_to_uint64_scalbn(a, s->float_rounding_mode, 0, s);
}
uint16_t bfloat16_to_uint16_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_uint16_scalbn(a, float_round_to_zero, 0, s);
}
uint32_t bfloat16_to_uint32_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_uint32_scalbn(a, float_round_to_zero, 0, s);
}
uint64_t bfloat16_to_uint64_round_to_zero(bfloat16 a, float_status *s)
{
return bfloat16_to_uint64_scalbn(a, float_round_to_zero, 0, s);
}
/*
* Signed integer to floating-point conversions
*/
float16 int64_to_float16_scalbn(int64_t a, int scale, float_status *status)
{
FloatParts64 p;
parts_sint_to_float(&p, a, scale, status);
return float16_round_pack_canonical(&p, status);
}
float16 int32_to_float16_scalbn(int32_t a, int scale, float_status *status)
{
return int64_to_float16_scalbn(a, scale, status);
}
float16 int16_to_float16_scalbn(int16_t a, int scale, float_status *status)
{
return int64_to_float16_scalbn(a, scale, status);
}
float16 int64_to_float16(int64_t a, float_status *status)
{
return int64_to_float16_scalbn(a, 0, status);
}
float16 int32_to_float16(int32_t a, float_status *status)
{
return int64_to_float16_scalbn(a, 0, status);
}
float16 int16_to_float16(int16_t a, float_status *status)
{
return int64_to_float16_scalbn(a, 0, status);
}
float16 int8_to_float16(int8_t a, float_status *status)
{
return int64_to_float16_scalbn(a, 0, status);
}
float32 int64_to_float32_scalbn(int64_t a, int scale, float_status *status)
{
FloatParts64 p;
/* Without scaling, there are no overflow concerns. */
if (likely(scale == 0) && can_use_fpu(status)) {
union_float32 ur;
ur.h = a;
return ur.s;
}
parts64_sint_to_float(&p, a, scale, status);
return float32_round_pack_canonical(&p, status);
}
float32 int32_to_float32_scalbn(int32_t a, int scale, float_status *status)
{
return int64_to_float32_scalbn(a, scale, status);
}
float32 int16_to_float32_scalbn(int16_t a, int scale, float_status *status)
{
return int64_to_float32_scalbn(a, scale, status);
}
float32 int64_to_float32(int64_t a, float_status *status)
{
return int64_to_float32_scalbn(a, 0, status);
}
float32 int32_to_float32(int32_t a, float_status *status)
{
return int64_to_float32_scalbn(a, 0, status);
}
float32 int16_to_float32(int16_t a, float_status *status)
{
return int64_to_float32_scalbn(a, 0, status);
}
float64 int64_to_float64_scalbn(int64_t a, int scale, float_status *status)
{
FloatParts64 p;
/* Without scaling, there are no overflow concerns. */
if (likely(scale == 0) && can_use_fpu(status)) {
union_float64 ur;
ur.h = a;
return ur.s;
}
parts_sint_to_float(&p, a, scale, status);
return float64_round_pack_canonical(&p, status);
}
float64 int32_to_float64_scalbn(int32_t a, int scale, float_status *status)
{
return int64_to_float64_scalbn(a, scale, status);
}
float64 int16_to_float64_scalbn(int16_t a, int scale, float_status *status)
{
return int64_to_float64_scalbn(a, scale, status);
}
float64 int64_to_float64(int64_t a, float_status *status)
{
return int64_to_float64_scalbn(a, 0, status);
}
float64 int32_to_float64(int32_t a, float_status *status)
{
return int64_to_float64_scalbn(a, 0, status);
}
float64 int16_to_float64(int16_t a, float_status *status)
{
return int64_to_float64_scalbn(a, 0, status);
}
bfloat16 int64_to_bfloat16_scalbn(int64_t a, int scale, float_status *status)
{
FloatParts64 p;
parts_sint_to_float(&p, a, scale, status);
return bfloat16_round_pack_canonical(&p, status);
}
bfloat16 int32_to_bfloat16_scalbn(int32_t a, int scale, float_status *status)
{
return int64_to_bfloat16_scalbn(a, scale, status);
}
bfloat16 int16_to_bfloat16_scalbn(int16_t a, int scale, float_status *status)
{
return int64_to_bfloat16_scalbn(a, scale, status);
}
bfloat16 int64_to_bfloat16(int64_t a, float_status *status)
{
return int64_to_bfloat16_scalbn(a, 0, status);
}
bfloat16 int32_to_bfloat16(int32_t a, float_status *status)
{
return int64_to_bfloat16_scalbn(a, 0, status);
}
bfloat16 int16_to_bfloat16(int16_t a, float_status *status)
{
return int64_to_bfloat16_scalbn(a, 0, status);
}
float128 int128_to_float128(Int128 a, float_status *status)
{
FloatParts128 p = { };
int shift;
if (int128_nz(a)) {
p.cls = float_class_normal;
if (!int128_nonneg(a)) {
p.sign = true;
a = int128_neg(a);
}
shift = clz64(int128_gethi(a));
if (shift == 64) {
shift += clz64(int128_getlo(a));
}
p.exp = 127 - shift;
a = int128_lshift(a, shift);
p.frac_hi = int128_gethi(a);
p.frac_lo = int128_getlo(a);
} else {
p.cls = float_class_zero;
}
return float128_round_pack_canonical(&p, status);
}
float128 int64_to_float128(int64_t a, float_status *status)
{
FloatParts128 p;
parts_sint_to_float(&p, a, 0, status);
return float128_round_pack_canonical(&p, status);
}
float128 int32_to_float128(int32_t a, float_status *status)
{
return int64_to_float128(a, status);
}
floatx80 int64_to_floatx80(int64_t a, float_status *status)
{
FloatParts128 p;
parts_sint_to_float(&p, a, 0, status);
return floatx80_round_pack_canonical(&p, status);
}
floatx80 int32_to_floatx80(int32_t a, float_status *status)
{
return int64_to_floatx80(a, status);
}
/*
* Unsigned Integer to floating-point conversions
*/
float16 uint64_to_float16_scalbn(uint64_t a, int scale, float_status *status)
{
FloatParts64 p;
parts_uint_to_float(&p, a, scale, status);
return float16_round_pack_canonical(&p, status);
}
float16 uint32_to_float16_scalbn(uint32_t a, int scale, float_status *status)
{
return uint64_to_float16_scalbn(a, scale, status);
}
float16 uint16_to_float16_scalbn(uint16_t a, int scale, float_status *status)
{
return uint64_to_float16_scalbn(a, scale, status);
}
float16 uint64_to_float16(uint64_t a, float_status *status)
{
return uint64_to_float16_scalbn(a, 0, status);
}
float16 uint32_to_float16(uint32_t a, float_status *status)
{
return uint64_to_float16_scalbn(a, 0, status);
}
float16 uint16_to_float16(uint16_t a, float_status *status)
{
return uint64_to_float16_scalbn(a, 0, status);
}
float16 uint8_to_float16(uint8_t a, float_status *status)
{
return uint64_to_float16_scalbn(a, 0, status);
}
float32 uint64_to_float32_scalbn(uint64_t a, int scale, float_status *status)
{
FloatParts64 p;
/* Without scaling, there are no overflow concerns. */
if (likely(scale == 0) && can_use_fpu(status)) {
union_float32 ur;
ur.h = a;
return ur.s;
}
parts_uint_to_float(&p, a, scale, status);
return float32_round_pack_canonical(&p, status);
}
float32 uint32_to_float32_scalbn(uint32_t a, int scale, float_status *status)
{
return uint64_to_float32_scalbn(a, scale, status);
}
float32 uint16_to_float32_scalbn(uint16_t a, int scale, float_status *status)
{
return uint64_to_float32_scalbn(a, scale, status);
}
float32 uint64_to_float32(uint64_t a, float_status *status)
{
return uint64_to_float32_scalbn(a, 0, status);
}
float32 uint32_to_float32(uint32_t a, float_status *status)
{
return uint64_to_float32_scalbn(a, 0, status);
}
float32 uint16_to_float32(uint16_t a, float_status *status)
{
return uint64_to_float32_scalbn(a, 0, status);
}
float64 uint64_to_float64_scalbn(uint64_t a, int scale, float_status *status)
{
FloatParts64 p;
/* Without scaling, there are no overflow concerns. */
if (likely(scale == 0) && can_use_fpu(status)) {
union_float64 ur;
ur.h = a;
return ur.s;
}
parts_uint_to_float(&p, a, scale, status);
return float64_round_pack_canonical(&p, status);
}
float64 uint32_to_float64_scalbn(uint32_t a, int scale, float_status *status)
{
return uint64_to_float64_scalbn(a, scale, status);
}
float64 uint16_to_float64_scalbn(uint16_t a, int scale, float_status *status)
{
return uint64_to_float64_scalbn(a, scale, status);
}
float64 uint64_to_float64(uint64_t a, float_status *status)
{
return uint64_to_float64_scalbn(a, 0, status);
}
float64 uint32_to_float64(uint32_t a, float_status *status)
{
return uint64_to_float64_scalbn(a, 0, status);
}
float64 uint16_to_float64(uint16_t a, float_status *status)
{
return uint64_to_float64_scalbn(a, 0, status);
}
bfloat16 uint64_to_bfloat16_scalbn(uint64_t a, int scale, float_status *status)
{
FloatParts64 p;
parts_uint_to_float(&p, a, scale, status);
return bfloat16_round_pack_canonical(&p, status);
}
bfloat16 uint32_to_bfloat16_scalbn(uint32_t a, int scale, float_status *status)
{
return uint64_to_bfloat16_scalbn(a, scale, status);
}
bfloat16 uint16_to_bfloat16_scalbn(uint16_t a, int scale, float_status *status)
{
return uint64_to_bfloat16_scalbn(a, scale, status);
}
bfloat16 uint64_to_bfloat16(uint64_t a, float_status *status)
{
return uint64_to_bfloat16_scalbn(a, 0, status);
}
bfloat16 uint32_to_bfloat16(uint32_t a, float_status *status)
{
return uint64_to_bfloat16_scalbn(a, 0, status);
}
bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status)
{
return uint64_to_bfloat16_scalbn(a, 0, status);
}
float128 uint64_to_float128(uint64_t a, float_status *status)
{
FloatParts128 p;
parts_uint_to_float(&p, a, 0, status);
return float128_round_pack_canonical(&p, status);
}
float128 uint128_to_float128(Int128 a, float_status *status)
{
FloatParts128 p = { };
int shift;
if (int128_nz(a)) {
p.cls = float_class_normal;
shift = clz64(int128_gethi(a));
if (shift == 64) {
shift += clz64(int128_getlo(a));
}
p.exp = 127 - shift;
a = int128_lshift(a, shift);
p.frac_hi = int128_gethi(a);
p.frac_lo = int128_getlo(a);
} else {
p.cls = float_class_zero;
}
return float128_round_pack_canonical(&p, status);
}
/*
* Minimum and maximum
*/
static float16 float16_minmax(float16 a, float16 b, float_status *s, int flags)
{
FloatParts64 pa, pb, *pr;
float16_unpack_canonical(&pa, a, s);
float16_unpack_canonical(&pb, b, s);
pr = parts_minmax(&pa, &pb, s, flags);
return float16_round_pack_canonical(pr, s);
}
static bfloat16 bfloat16_minmax(bfloat16 a, bfloat16 b,
float_status *s, int flags)
{
FloatParts64 pa, pb, *pr;
bfloat16_unpack_canonical(&pa, a, s);
bfloat16_unpack_canonical(&pb, b, s);
pr = parts_minmax(&pa, &pb, s, flags);
return bfloat16_round_pack_canonical(pr, s);
}
static float32 float32_minmax(float32 a, float32 b, float_status *s, int flags)
{
FloatParts64 pa, pb, *pr;
float32_unpack_canonical(&pa, a, s);
float32_unpack_canonical(&pb, b, s);
pr = parts_minmax(&pa, &pb, s, flags);
return float32_round_pack_canonical(pr, s);
}
static float64 float64_minmax(float64 a, float64 b, float_status *s, int flags)
{
FloatParts64 pa, pb, *pr;
float64_unpack_canonical(&pa, a, s);
float64_unpack_canonical(&pb, b, s);
pr = parts_minmax(&pa, &pb, s, flags);
return float64_round_pack_canonical(pr, s);
}
static float128 float128_minmax(float128 a, float128 b,
float_status *s, int flags)
{
FloatParts128 pa, pb, *pr;
float128_unpack_canonical(&pa, a, s);
float128_unpack_canonical(&pb, b, s);
pr = parts_minmax(&pa, &pb, s, flags);
return float128_round_pack_canonical(pr, s);
}
#define MINMAX_1(type, name, flags) \
type type##_##name(type a, type b, float_status *s) \
{ return type##_minmax(a, b, s, flags); }
#define MINMAX_2(type) \
MINMAX_1(type, max, 0) \
MINMAX_1(type, maxnum, minmax_isnum) \
MINMAX_1(type, maxnummag, minmax_isnum | minmax_ismag) \
MINMAX_1(type, maximum_number, minmax_isnumber) \
MINMAX_1(type, min, minmax_ismin) \
MINMAX_1(type, minnum, minmax_ismin | minmax_isnum) \
MINMAX_1(type, minnummag, minmax_ismin | minmax_isnum | minmax_ismag) \
MINMAX_1(type, minimum_number, minmax_ismin | minmax_isnumber) \
MINMAX_2(float16)
MINMAX_2(bfloat16)
MINMAX_2(float32)
MINMAX_2(float64)
MINMAX_2(float128)
#undef MINMAX_1
#undef MINMAX_2
/*
* Floating point compare
*/
static FloatRelation QEMU_FLATTEN
float16_do_compare(float16 a, float16 b, float_status *s, bool is_quiet)
{
FloatParts64 pa, pb;
float16_unpack_canonical(&pa, a, s);
float16_unpack_canonical(&pb, b, s);
return parts_compare(&pa, &pb, s, is_quiet);
}
FloatRelation float16_compare(float16 a, float16 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float16_do_compare(a, b, s, false);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
FloatRelation float16_compare_quiet(float16 a, float16 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float16_do_compare(a, b, s, true);
}
static FloatRelation QEMU_SOFTFLOAT_ATTR
float32_do_compare(float32 a, float32 b, float_status *s, bool is_quiet)
{
FloatParts64 pa, pb;
float32_unpack_canonical(&pa, a, s);
float32_unpack_canonical(&pb, b, s);
return parts_compare(&pa, &pb, s, is_quiet);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
static FloatRelation QEMU_FLATTEN
float32_hs_compare(float32 xa, float32 xb, float_status *s, bool is_quiet)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
union_float32 ua, ub;
ua.s = xa;
ub.s = xb;
if (QEMU_NO_HARDFLOAT) {
goto soft;
}
float32_input_flush2(&ua.s, &ub.s, s);
if (isgreaterequal(ua.h, ub.h)) {
if (isgreater(ua.h, ub.h)) {
return float_relation_greater;
}
return float_relation_equal;
}
if (likely(isless(ua.h, ub.h))) {
return float_relation_less;
}
/*
* The only condition remaining is unordered.
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
* Fall through to set flags.
*/
soft:
return float32_do_compare(ua.s, ub.s, s, is_quiet);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
FloatRelation float32_compare(float32 a, float32 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float32_hs_compare(a, b, s, false);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
FloatRelation float32_compare_quiet(float32 a, float32 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float32_hs_compare(a, b, s, true);
}
static FloatRelation QEMU_SOFTFLOAT_ATTR
float64_do_compare(float64 a, float64 b, float_status *s, bool is_quiet)
{
FloatParts64 pa, pb;
float64_unpack_canonical(&pa, a, s);
float64_unpack_canonical(&pb, b, s);
return parts_compare(&pa, &pb, s, is_quiet);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
static FloatRelation QEMU_FLATTEN
float64_hs_compare(float64 xa, float64 xb, float_status *s, bool is_quiet)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
union_float64 ua, ub;
ua.s = xa;
ub.s = xb;
if (QEMU_NO_HARDFLOAT) {
goto soft;
}
float64_input_flush2(&ua.s, &ub.s, s);
if (isgreaterequal(ua.h, ub.h)) {
if (isgreater(ua.h, ub.h)) {
return float_relation_greater;
}
return float_relation_equal;
}
if (likely(isless(ua.h, ub.h))) {
return float_relation_less;
}
/*
* The only condition remaining is unordered.
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
* Fall through to set flags.
*/
soft:
return float64_do_compare(ua.s, ub.s, s, is_quiet);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
FloatRelation float64_compare(float64 a, float64 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float64_hs_compare(a, b, s, false);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
FloatRelation float64_compare_quiet(float64 a, float64 b, float_status *s)
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
{
return float64_hs_compare(a, b, s, true);
hardfloat: implement float32/64 comparison Performance results for fp-bench: Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz - before: cmp-single: 110.98 MFlops cmp-double: 107.12 MFlops - after: cmp-single: 506.28 MFlops cmp-double: 524.77 MFlops Note that flattening both eq and eq_signaling versions would give us extra performance (695v506, 615v524 Mflops for single/double, respectively) but this would emit two essentially identical functions for each eq/signaling pair, which is a waste. Aggregate performance improvement for the last few patches: [ all charts in png: https://imgur.com/a/4yV8p ] 1. Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz qemu-aarch64 NBench score; higher is better Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 16 +-+-----------+-------------+----===-------+---===-------+-----------+-+ 14 +-+..........................@@@&&.=.......@@@&&.=...................+-+ 12 +-+..........................@.@.&.=.......@.@.&.=.....+befor=== +-+ 10 +-+..........................@.@.&.=.......@.@.&.=.....+ad@@&& = +-+ 8 +-+.......................$$$%.@.&.=.......@.@.&.=.....+ @@u& = +-+ 6 +-+............@@@&&=+***##.$%.@.&.=***##$$%+@.&.=..###$$%%@i& = +-+ 4 +-+.......###$%%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=+**.#+$ +@m& = +-+ 2 +-+.....***.#$.%.@.&=.*.*.#.$%.@.&.=*.*.#.$%.@.&.=.**.#+$+sqr& = +-+ 0 +-+-----***##$%%@@&&=-***##$$%@@&&==***##$$%@@&&==-**##$$%+cmp==-----+-+ FOURIER NEURAL NELU DECOMPOSITION gmean qemu-aarch64 SPEC06fp (test set) speedup over QEMU 4c2c1015905 Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz error bars: 95% confidence interval 4.5 +-+---+-----+----+-----+-----+-&---+-----+----+-----+-----+-----+----+-----+-----+-----+-----+----+-----+---+-+ 4 +-+..........................+@@+...........................................................................+-+ 3.5 +-+..............%%@&.........@@..............%%@&............................................+++dsub +-+ 2.5 +-+....&&+.......%%@&.......+%%@..+%%&+..@@&+.%%@&....................................+%%&+.+%@&++%%@& +-+ 2 +-+..+%%&..+%@&+.%%@&...+++..%%@...%%&.+$$@&..%%@&..%%@&.......+%%&+.%%@&+......+%%@&.+%%&++$$@&++d%@& %%@&+-+ 1.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+f%@&**$%@&+-+ 0.5 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&+sqr@&**$%@&+-+ 0 +-+**#$%&**#$@&**#%@&**$%@**#$%@**#$%&**#$@&**$%@&*#$%@**#$%@**#$%&**#%@&**$%@&*#$%@**#$%&**#$@&*+cmp&**$%@&+-+ 410.bw416.gam433.434.z435.436.cac437.lesli444.447.de450.so453454.ca459.GemsF465.tont470.lb4482.sphinxgeomean 2. Host: ARM Aarch64 A57 @ 2.4GHz qemu-aarch64 NBench score; higher is better Host: Applied Micro X-Gene, Aarch64 A57 @ 2.4 GHz 5 +-+-----------+-------------+-------------+-------------+-----------+-+ 4.5 +-+........................................@@@&==...................+-+ 3 4 +-+..........................@@@&==........@.@&.=.....+before +-+ 3 +-+..........................@.@&.=........@.@&.=.....+ad@@@&== +-+ 2.5 +-+.....................##$$%%.@&.=........@.@&.=.....+ @m@& = +-+ 2 +-+............@@@&==.***#.$.%.@&.=.***#$$%%.@&.=.***#$$%%d@& = +-+ 1.5 +-+.....***#$$%%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$ +f@& = +-+ 0.5 +-+.....*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#.$.%.@&.=.*.*#+$+sqr& = +-+ 0 +-+-----***#$$%%@@&==-***#$$%%@@&==-***#$$%%@@&==-***#$$%+cmp==-----+-+ FOURIER NEURAL NLU DECOMPOSITION gmean Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
2018-03-17 09:14:53 +03:00
}
static FloatRelation QEMU_FLATTEN
bfloat16_do_compare(bfloat16 a, bfloat16 b, float_status *s, bool is_quiet)
{
FloatParts64 pa, pb;
bfloat16_unpack_canonical(&pa, a, s);
bfloat16_unpack_canonical(&pb, b, s);
return parts_compare(&pa, &pb, s, is_quiet);
}
FloatRelation bfloat16_compare(bfloat16 a, bfloat16 b, float_status *s)
{
return bfloat16_do_compare(a, b, s, false);
}
FloatRelation bfloat16_compare_quiet(bfloat16 a, bfloat16 b, float_status *s)
{
return bfloat16_do_compare(a, b, s, true);
}
static FloatRelation QEMU_FLATTEN
float128_do_compare(float128 a, float128 b, float_status *s, bool is_quiet)
{
FloatParts128 pa, pb;
float128_unpack_canonical(&pa, a, s);
float128_unpack_canonical(&pb, b, s);
return parts_compare(&pa, &pb, s, is_quiet);
}
FloatRelation float128_compare(float128 a, float128 b, float_status *s)
{
return float128_do_compare(a, b, s, false);
}
FloatRelation float128_compare_quiet(float128 a, float128 b, float_status *s)
{
return float128_do_compare(a, b, s, true);
}
static FloatRelation QEMU_FLATTEN
floatx80_do_compare(floatx80 a, floatx80 b, float_status *s, bool is_quiet)
{
FloatParts128 pa, pb;
if (!floatx80_unpack_canonical(&pa, a, s) ||
!floatx80_unpack_canonical(&pb, b, s)) {
return float_relation_unordered;
}
return parts_compare(&pa, &pb, s, is_quiet);
}
FloatRelation floatx80_compare(floatx80 a, floatx80 b, float_status *s)
{
return floatx80_do_compare(a, b, s, false);
}
FloatRelation floatx80_compare_quiet(floatx80 a, floatx80 b, float_status *s)
{
return floatx80_do_compare(a, b, s, true);
}
/*
* Scale by 2**N
*/
float16 float16_scalbn(float16 a, int n, float_status *status)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, status);
parts_scalbn(&p, n, status);
return float16_round_pack_canonical(&p, status);
}
float32 float32_scalbn(float32 a, int n, float_status *status)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, status);
parts_scalbn(&p, n, status);
return float32_round_pack_canonical(&p, status);
}
float64 float64_scalbn(float64 a, int n, float_status *status)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, status);
parts_scalbn(&p, n, status);
return float64_round_pack_canonical(&p, status);
}
bfloat16 bfloat16_scalbn(bfloat16 a, int n, float_status *status)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, status);
parts_scalbn(&p, n, status);
return bfloat16_round_pack_canonical(&p, status);
}
float128 float128_scalbn(float128 a, int n, float_status *status)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, status);
parts_scalbn(&p, n, status);
return float128_round_pack_canonical(&p, status);
}
floatx80 floatx80_scalbn(floatx80 a, int n, float_status *status)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, status)) {
return floatx80_default_nan(status);
}
parts_scalbn(&p, n, status);
return floatx80_round_pack_canonical(&p, status);
}
/*
* Square Root
*/
float16 QEMU_FLATTEN float16_sqrt(float16 a, float_status *status)
{
FloatParts64 p;
float16_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &float16_params);
return float16_round_pack_canonical(&p, status);
}
static float32 QEMU_SOFTFLOAT_ATTR
soft_f32_sqrt(float32 a, float_status *status)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &float32_params);
return float32_round_pack_canonical(&p, status);
}
static float64 QEMU_SOFTFLOAT_ATTR
soft_f64_sqrt(float64 a, float_status *status)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &float64_params);
return float64_round_pack_canonical(&p, status);
}
float32 QEMU_FLATTEN float32_sqrt(float32 xa, float_status *s)
{
union_float32 ua, ur;
ua.s = xa;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
float32_input_flush1(&ua.s, s);
if (QEMU_HARDFLOAT_1F32_USE_FP) {
if (unlikely(!(fpclassify(ua.h) == FP_NORMAL ||
fpclassify(ua.h) == FP_ZERO) ||
signbit(ua.h))) {
goto soft;
}
} else if (unlikely(!float32_is_zero_or_normal(ua.s) ||
float32_is_neg(ua.s))) {
goto soft;
}
ur.h = sqrtf(ua.h);
return ur.s;
soft:
return soft_f32_sqrt(ua.s, s);
}
float64 QEMU_FLATTEN float64_sqrt(float64 xa, float_status *s)
{
union_float64 ua, ur;
ua.s = xa;
if (unlikely(!can_use_fpu(s))) {
goto soft;
}
float64_input_flush1(&ua.s, s);
if (QEMU_HARDFLOAT_1F64_USE_FP) {
if (unlikely(!(fpclassify(ua.h) == FP_NORMAL ||
fpclassify(ua.h) == FP_ZERO) ||
signbit(ua.h))) {
goto soft;
}
} else if (unlikely(!float64_is_zero_or_normal(ua.s) ||
float64_is_neg(ua.s))) {
goto soft;
}
ur.h = sqrt(ua.h);
return ur.s;
soft:
return soft_f64_sqrt(ua.s, s);
}
float64 float64r32_sqrt(float64 a, float_status *status)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &float64_params);
return float64r32_round_pack_canonical(&p, status);
}
bfloat16 QEMU_FLATTEN bfloat16_sqrt(bfloat16 a, float_status *status)
{
FloatParts64 p;
bfloat16_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &bfloat16_params);
return bfloat16_round_pack_canonical(&p, status);
}
float128 QEMU_FLATTEN float128_sqrt(float128 a, float_status *status)
{
FloatParts128 p;
float128_unpack_canonical(&p, a, status);
parts_sqrt(&p, status, &float128_params);
return float128_round_pack_canonical(&p, status);
}
floatx80 floatx80_sqrt(floatx80 a, float_status *s)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, s)) {
return floatx80_default_nan(s);
}
parts_sqrt(&p, s, &floatx80_params[s->floatx80_rounding_precision]);
return floatx80_round_pack_canonical(&p, s);
}
/*
* log2
*/
float32 float32_log2(float32 a, float_status *status)
{
FloatParts64 p;
float32_unpack_canonical(&p, a, status);
parts_log2(&p, status, &float32_params);
return float32_round_pack_canonical(&p, status);
}
float64 float64_log2(float64 a, float_status *status)
{
FloatParts64 p;
float64_unpack_canonical(&p, a, status);
parts_log2(&p, status, &float64_params);
return float64_round_pack_canonical(&p, status);
}
/*----------------------------------------------------------------------------
| The pattern for a default generated NaN.
*----------------------------------------------------------------------------*/
float16 float16_default_nan(float_status *status)
{
FloatParts64 p;
parts_default_nan(&p, status);
p.frac >>= float16_params.frac_shift;
return float16_pack_raw(&p);
}
float32 float32_default_nan(float_status *status)
{
FloatParts64 p;
parts_default_nan(&p, status);
p.frac >>= float32_params.frac_shift;
return float32_pack_raw(&p);
}
float64 float64_default_nan(float_status *status)
{
FloatParts64 p;
parts_default_nan(&p, status);
p.frac >>= float64_params.frac_shift;
return float64_pack_raw(&p);
}
float128 float128_default_nan(float_status *status)
{
FloatParts128 p;
parts_default_nan(&p, status);
frac_shr(&p, float128_params.frac_shift);
return float128_pack_raw(&p);
}
bfloat16 bfloat16_default_nan(float_status *status)
{
FloatParts64 p;
parts_default_nan(&p, status);
p.frac >>= bfloat16_params.frac_shift;
return bfloat16_pack_raw(&p);
}
/*----------------------------------------------------------------------------
| Returns a quiet NaN from a signalling NaN for the floating point value `a'.
*----------------------------------------------------------------------------*/
float16 float16_silence_nan(float16 a, float_status *status)
{
FloatParts64 p;
float16_unpack_raw(&p, a);
p.frac <<= float16_params.frac_shift;
parts_silence_nan(&p, status);
p.frac >>= float16_params.frac_shift;
return float16_pack_raw(&p);
}
float32 float32_silence_nan(float32 a, float_status *status)
{
FloatParts64 p;
float32_unpack_raw(&p, a);
p.frac <<= float32_params.frac_shift;
parts_silence_nan(&p, status);
p.frac >>= float32_params.frac_shift;
return float32_pack_raw(&p);
}
float64 float64_silence_nan(float64 a, float_status *status)
{
FloatParts64 p;
float64_unpack_raw(&p, a);
p.frac <<= float64_params.frac_shift;
parts_silence_nan(&p, status);
p.frac >>= float64_params.frac_shift;
return float64_pack_raw(&p);
}
bfloat16 bfloat16_silence_nan(bfloat16 a, float_status *status)
{
FloatParts64 p;
bfloat16_unpack_raw(&p, a);
p.frac <<= bfloat16_params.frac_shift;
parts_silence_nan(&p, status);
p.frac >>= bfloat16_params.frac_shift;
return bfloat16_pack_raw(&p);
}
float128 float128_silence_nan(float128 a, float_status *status)
{
FloatParts128 p;
float128_unpack_raw(&p, a);
frac_shl(&p, float128_params.frac_shift);
parts_silence_nan(&p, status);
frac_shr(&p, float128_params.frac_shift);
return float128_pack_raw(&p);
}
/*----------------------------------------------------------------------------
| If `a' is denormal and we are in flush-to-zero mode then set the
| input-denormal exception and return zero. Otherwise just return the value.
*----------------------------------------------------------------------------*/
static bool parts_squash_denormal(FloatParts64 p, float_status *status)
{
if (p.exp == 0 && p.frac != 0) {
float_raise(float_flag_input_denormal, status);
return true;
}
return false;
}
float16 float16_squash_input_denormal(float16 a, float_status *status)
{
if (status->flush_inputs_to_zero) {
FloatParts64 p;
float16_unpack_raw(&p, a);
if (parts_squash_denormal(p, status)) {
return float16_set_sign(float16_zero, p.sign);
}
}
return a;
}
float32 float32_squash_input_denormal(float32 a, float_status *status)
{
if (status->flush_inputs_to_zero) {
FloatParts64 p;
float32_unpack_raw(&p, a);
if (parts_squash_denormal(p, status)) {
return float32_set_sign(float32_zero, p.sign);
}
}
return a;
}
float64 float64_squash_input_denormal(float64 a, float_status *status)
{
if (status->flush_inputs_to_zero) {
FloatParts64 p;
float64_unpack_raw(&p, a);
if (parts_squash_denormal(p, status)) {
return float64_set_sign(float64_zero, p.sign);
}
}
return a;
}
bfloat16 bfloat16_squash_input_denormal(bfloat16 a, float_status *status)
{
if (status->flush_inputs_to_zero) {
FloatParts64 p;
bfloat16_unpack_raw(&p, a);
if (parts_squash_denormal(p, status)) {
return bfloat16_set_sign(bfloat16_zero, p.sign);
}
}
return a;
}
/*----------------------------------------------------------------------------
| Normalizes the subnormal extended double-precision floating-point value
| represented by the denormalized significand `aSig'. The normalized exponent
| and significand are stored at the locations pointed to by `zExpPtr' and
| `zSigPtr', respectively.
*----------------------------------------------------------------------------*/
void normalizeFloatx80Subnormal(uint64_t aSig, int32_t *zExpPtr,
uint64_t *zSigPtr)
{
int8_t shiftCount;
shiftCount = clz64(aSig);
*zSigPtr = aSig<<shiftCount;
*zExpPtr = 1 - shiftCount;
}
/*----------------------------------------------------------------------------
| Takes an abstract floating-point value having sign `zSign', exponent `zExp',
| and extended significand formed by the concatenation of `zSig0' and `zSig1',
| and returns the proper extended double-precision floating-point value
| corresponding to the abstract input. Ordinarily, the abstract value is
| rounded and packed into the extended double-precision format, with the
| inexact exception raised if the abstract input cannot be represented
| exactly. However, if the abstract value is too large, the overflow and
| inexact exceptions are raised and an infinity or maximal finite value is
| returned. If the abstract value is too small, the input value is rounded to
| a subnormal number, and the underflow and inexact exceptions are raised if
| the abstract input cannot be represented exactly as a subnormal extended
| double-precision floating-point number.
| If `roundingPrecision' is floatx80_precision_s or floatx80_precision_d,
| the result is rounded to the same number of bits as single or double
| precision, respectively. Otherwise, the result is rounded to the full
| precision of the extended double-precision format.
| The input significand must be normalized or smaller. If the input
| significand is not normalized, `zExp' must be 0; in that case, the result
| returned is a subnormal number, and it must not require rounding. The
| handling of underflow and overflow follows the IEC/IEEE Standard for Binary
| Floating-Point Arithmetic.
*----------------------------------------------------------------------------*/
floatx80 roundAndPackFloatx80(FloatX80RoundPrec roundingPrecision, bool zSign,
int32_t zExp, uint64_t zSig0, uint64_t zSig1,
float_status *status)
{
FloatRoundMode roundingMode;
bool roundNearestEven, increment, isTiny;
int64_t roundIncrement, roundMask, roundBits;
roundingMode = status->float_rounding_mode;
roundNearestEven = ( roundingMode == float_round_nearest_even );
switch (roundingPrecision) {
case floatx80_precision_x:
goto precision80;
case floatx80_precision_d:
roundIncrement = UINT64_C(0x0000000000000400);
roundMask = UINT64_C(0x00000000000007FF);
break;
case floatx80_precision_s:
roundIncrement = UINT64_C(0x0000008000000000);
roundMask = UINT64_C(0x000000FFFFFFFFFF);
break;
default:
g_assert_not_reached();
}
zSig0 |= ( zSig1 != 0 );
switch (roundingMode) {
case float_round_nearest_even:
case float_round_ties_away:
break;
case float_round_to_zero:
roundIncrement = 0;
break;
case float_round_up:
roundIncrement = zSign ? 0 : roundMask;
break;
case float_round_down:
roundIncrement = zSign ? roundMask : 0;
break;
default:
abort();
}
roundBits = zSig0 & roundMask;
if ( 0x7FFD <= (uint32_t) ( zExp - 1 ) ) {
if ( ( 0x7FFE < zExp )
|| ( ( zExp == 0x7FFE ) && ( zSig0 + roundIncrement < zSig0 ) )
) {
goto overflow;
}
if ( zExp <= 0 ) {
if (status->flush_to_zero) {
float_raise(float_flag_output_denormal, status);
return packFloatx80(zSign, 0, 0);
}
isTiny = status->tininess_before_rounding
|| (zExp < 0 )
|| (zSig0 <= zSig0 + roundIncrement);
shift64RightJamming( zSig0, 1 - zExp, &zSig0 );
zExp = 0;
roundBits = zSig0 & roundMask;
if (isTiny && roundBits) {
float_raise(float_flag_underflow, status);
}
if (roundBits) {
float_raise(float_flag_inexact, status);
}
zSig0 += roundIncrement;
if ( (int64_t) zSig0 < 0 ) zExp = 1;
roundIncrement = roundMask + 1;
if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) {
roundMask |= roundIncrement;
}
zSig0 &= ~ roundMask;
return packFloatx80( zSign, zExp, zSig0 );
}
}
if (roundBits) {
float_raise(float_flag_inexact, status);
}
zSig0 += roundIncrement;
if ( zSig0 < roundIncrement ) {
++zExp;
zSig0 = UINT64_C(0x8000000000000000);
}
roundIncrement = roundMask + 1;
if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) {
roundMask |= roundIncrement;
}
zSig0 &= ~ roundMask;
if ( zSig0 == 0 ) zExp = 0;
return packFloatx80( zSign, zExp, zSig0 );
precision80:
switch (roundingMode) {
case float_round_nearest_even:
case float_round_ties_away:
increment = ((int64_t)zSig1 < 0);
break;
case float_round_to_zero:
increment = 0;
break;
case float_round_up:
increment = !zSign && zSig1;
break;
case float_round_down:
increment = zSign && zSig1;
break;
default:
abort();
}
if ( 0x7FFD <= (uint32_t) ( zExp - 1 ) ) {
if ( ( 0x7FFE < zExp )
|| ( ( zExp == 0x7FFE )
&& ( zSig0 == UINT64_C(0xFFFFFFFFFFFFFFFF) )
&& increment
)
) {
roundMask = 0;
overflow:
float_raise(float_flag_overflow | float_flag_inexact, status);
if ( ( roundingMode == float_round_to_zero )
|| ( zSign && ( roundingMode == float_round_up ) )
|| ( ! zSign && ( roundingMode == float_round_down ) )
) {
return packFloatx80( zSign, 0x7FFE, ~ roundMask );
}
return packFloatx80(zSign,
floatx80_infinity_high,
floatx80_infinity_low);
}
if ( zExp <= 0 ) {
isTiny = status->tininess_before_rounding
|| (zExp < 0)
|| !increment
|| (zSig0 < UINT64_C(0xFFFFFFFFFFFFFFFF));
shift64ExtraRightJamming( zSig0, zSig1, 1 - zExp, &zSig0, &zSig1 );
zExp = 0;
if (isTiny && zSig1) {
float_raise(float_flag_underflow, status);
}
if (zSig1) {
float_raise(float_flag_inexact, status);
}
switch (roundingMode) {
case float_round_nearest_even:
case float_round_ties_away:
increment = ((int64_t)zSig1 < 0);
break;
case float_round_to_zero:
increment = 0;
break;
case float_round_up:
increment = !zSign && zSig1;
break;
case float_round_down:
increment = zSign && zSig1;
break;
default:
abort();
}
if ( increment ) {
++zSig0;
fpu/softfloat: Silence 'bitwise negation of boolean expression' warning When building with clang version 10.0.0-4ubuntu1, we get: CC lm32-softmmu/fpu/softfloat.o fpu/softfloat.c:3365:13: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] absZ &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ fpu/softfloat.c:3423:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] absZ0 &= ~ ( ( (uint64_t) ( absZ1<<1 ) == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ... fpu/softfloat.c:4273:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] zSig1 &= ~ ( ( zSig2 + zSig2 == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Fix by rewriting the fishy bitwise AND of two bools as an int. Suggested-by: Eric Blake <eblake@redhat.com> Buglink: https://bugs.launchpad.net/bugs/1881004 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Thomas Huth <thuth@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20200617201309.1640952-2-richard.henderson@linaro.org Message-Id: <20200528155420.9802-1-philmd@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2020-06-17 23:13:04 +03:00
if (!(zSig1 << 1) && roundNearestEven) {
zSig0 &= ~1;
}
if ( (int64_t) zSig0 < 0 ) zExp = 1;
}
return packFloatx80( zSign, zExp, zSig0 );
}
}
if (zSig1) {
float_raise(float_flag_inexact, status);
}
if ( increment ) {
++zSig0;
if ( zSig0 == 0 ) {
++zExp;
zSig0 = UINT64_C(0x8000000000000000);
}
else {
fpu/softfloat: Silence 'bitwise negation of boolean expression' warning When building with clang version 10.0.0-4ubuntu1, we get: CC lm32-softmmu/fpu/softfloat.o fpu/softfloat.c:3365:13: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] absZ &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ fpu/softfloat.c:3423:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] absZ0 &= ~ ( ( (uint64_t) ( absZ1<<1 ) == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ... fpu/softfloat.c:4273:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation] zSig1 &= ~ ( ( zSig2 + zSig2 == 0 ) & roundNearestEven ); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Fix by rewriting the fishy bitwise AND of two bools as an int. Suggested-by: Eric Blake <eblake@redhat.com> Buglink: https://bugs.launchpad.net/bugs/1881004 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Thomas Huth <thuth@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Message-id: 20200617201309.1640952-2-richard.henderson@linaro.org Message-Id: <20200528155420.9802-1-philmd@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
2020-06-17 23:13:04 +03:00
if (!(zSig1 << 1) && roundNearestEven) {
zSig0 &= ~1;
}
}
}
else {
if ( zSig0 == 0 ) zExp = 0;
}
return packFloatx80( zSign, zExp, zSig0 );
}
/*----------------------------------------------------------------------------
| Takes an abstract floating-point value having sign `zSign', exponent
| `zExp', and significand formed by the concatenation of `zSig0' and `zSig1',
| and returns the proper extended double-precision floating-point value
| corresponding to the abstract input. This routine is just like
| `roundAndPackFloatx80' except that the input significand does not have to be
| normalized.
*----------------------------------------------------------------------------*/
floatx80 normalizeRoundAndPackFloatx80(FloatX80RoundPrec roundingPrecision,
bool zSign, int32_t zExp,
uint64_t zSig0, uint64_t zSig1,
float_status *status)
{
int8_t shiftCount;
if ( zSig0 == 0 ) {
zSig0 = zSig1;
zSig1 = 0;
zExp -= 64;
}
shiftCount = clz64(zSig0);
shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 );
zExp -= shiftCount;
return roundAndPackFloatx80(roundingPrecision, zSign, zExp,
zSig0, zSig1, status);
}
/*----------------------------------------------------------------------------
| Returns the binary exponential of the single-precision floating-point value
| `a'. The operation is performed according to the IEC/IEEE Standard for
| Binary Floating-Point Arithmetic.
|
| Uses the following identities:
|
| 1. -------------------------------------------------------------------------
| x x*ln(2)
| 2 = e
|
| 2. -------------------------------------------------------------------------
| 2 3 4 5 n
| x x x x x x x
| e = 1 + --- + --- + --- + --- + --- + ... + --- + ...
| 1! 2! 3! 4! 5! n!
*----------------------------------------------------------------------------*/
static const float64 float32_exp2_coefficients[15] =
{
const_float64( 0x3ff0000000000000ll ), /* 1 */
const_float64( 0x3fe0000000000000ll ), /* 2 */
const_float64( 0x3fc5555555555555ll ), /* 3 */
const_float64( 0x3fa5555555555555ll ), /* 4 */
const_float64( 0x3f81111111111111ll ), /* 5 */
const_float64( 0x3f56c16c16c16c17ll ), /* 6 */
const_float64( 0x3f2a01a01a01a01all ), /* 7 */
const_float64( 0x3efa01a01a01a01all ), /* 8 */
const_float64( 0x3ec71de3a556c734ll ), /* 9 */
const_float64( 0x3e927e4fb7789f5cll ), /* 10 */
const_float64( 0x3e5ae64567f544e4ll ), /* 11 */
const_float64( 0x3e21eed8eff8d898ll ), /* 12 */
const_float64( 0x3de6124613a86d09ll ), /* 13 */
const_float64( 0x3da93974a8c07c9dll ), /* 14 */
const_float64( 0x3d6ae7f3e733b81fll ), /* 15 */
};
float32 float32_exp2(float32 a, float_status *status)
{
FloatParts64 xp, xnp, tp, rp;
int i;
float32_unpack_canonical(&xp, a, status);
if (unlikely(xp.cls != float_class_normal)) {
switch (xp.cls) {
case float_class_snan:
case float_class_qnan:
parts_return_nan(&xp, status);
return float32_round_pack_canonical(&xp, status);
case float_class_inf:
return xp.sign ? float32_zero : a;
case float_class_zero:
return float32_one;
default:
break;
}
g_assert_not_reached();
}
float_raise(float_flag_inexact, status);
float64_unpack_canonical(&tp, float64_ln2, status);
xp = *parts_mul(&xp, &tp, status);
xnp = xp;
float64_unpack_canonical(&rp, float64_one, status);
for (i = 0 ; i < 15 ; i++) {
float64_unpack_canonical(&tp, float32_exp2_coefficients[i], status);
rp = *parts_muladd(&tp, &xnp, &rp, 0, status);
xnp = *parts_mul(&xnp, &xp, status);
}
return float32_round_pack_canonical(&rp, status);
}
/*----------------------------------------------------------------------------
| Rounds the extended double-precision floating-point value `a'
| to the precision provided by floatx80_rounding_precision and returns the
| result as an extended double-precision floating-point value.
| The operation is performed according to the IEC/IEEE Standard for Binary
| Floating-Point Arithmetic.
*----------------------------------------------------------------------------*/
floatx80 floatx80_round(floatx80 a, float_status *status)
{
FloatParts128 p;
if (!floatx80_unpack_canonical(&p, a, status)) {
return floatx80_default_nan(status);
}
return floatx80_round_pack_canonical(&p, status);
}
static void __attribute__((constructor)) softfloat_init(void)
{
union_float64 ua, ub, uc, ur;
if (QEMU_NO_HARDFLOAT) {
return;
}
/*
* Test that the host's FMA is not obviously broken. For example,
* glibc < 2.23 can perform an incorrect FMA on certain hosts; see
* https://sourceware.org/bugzilla/show_bug.cgi?id=13304
*/
ua.s = 0x0020000000000001ULL;
ub.s = 0x3ca0000000000000ULL;
uc.s = 0x0020000000000000ULL;
ur.h = fma(ua.h, ub.h, uc.h);
if (ur.s != 0x0020000000000001ULL) {
force_soft_fma = true;
}
}