Age | Commit message (Collapse) | Author |
|
With Makefiles that have automatically generated dependencies, you
generated includes are set as dependencies of the Makefile, so that they
are built before everything else and they are available when first
building the .c files.
Alternatively you can use a fine-grained dependency, e.g.
target/arm/translate.o: target/arm/decode-neon-shared.inc.c
With Meson you have only one choice and it is a third option, namely
"build at the beginning of the corresponding target"; the way you
express it is to list the includes in the sources of that target.
The problem is that Meson decides if something is a source vs. a
generated include by looking at the extension: '.c', '.cc', '.m', '.C'
are sources, while everything else is considered an include---including
'.inc.c'.
Use '.c.inc' to avoid this, as it is consistent with our other convention
of using '.rst.inc' for included reStructuredText files. The editorconfig
file is adjusted.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
Signed-off-by: LIU Zhiwei <zhiwei_liu@c-sky.com>
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20200712234521.3972-2-zhiwei_liu@c-sky.com>
Message-Id: <20200713200415.26214-5-alex.bennee@linaro.org>
|
|
Both x87 and m68k need the low parts of the quotient for their
remainder operations. Arrange for floatx80_modrem to track those bits
and return them via a pointer.
The architectures using float32_rem and float64_rem do not appear to
need this information, so the *_rem interface is left unchanged and
the information returned only from floatx80_modrem. The logic used to
determine the low 7 bits of the quotient for m68k
(target/m68k/fpu_helper.c:make_quotient) appears completely bogus (it
looks at the result of converting the remainder to integer, the
quotient having been discarded by that point); this patch does not
change that, but the m68k maintainers may wish to do so.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <alpine.DEB.2.21.2006081656500.23637@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
The floatx80 remainder implementation unnecessarily sets the high bit
of bSig explicitly. By that point in the function, arguments that are
invalid, zero, infinity or NaN have already been handled and
subnormals have been through normalizeFloatx80Subnormal, so the high
bit will already be set. Remove the unnecessary code.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <alpine.DEB.2.21.2006081656220.23637@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
The floatx80 remainder implementation sometimes returns the numerator
unchanged when the denominator is sufficiently larger than the
numerator. But if the value to be returned unchanged is a
pseudo-denormal, that is incorrect. Fix it to normalize the numerator
in that case.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <alpine.DEB.2.21.2006081655520.23637@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
The floatx80 remainder implementation ignores the high bit of the
significand when checking whether an operand (numerator) with zero
exponent is zero. This means it mishandles a pseudo-denormal
representation of 0x1p-16382L by treating it as zero. Fix this by
checking the whole significand instead.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <alpine.DEB.2.21.2006081655180.23637@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
The m68k-specific softfloat code includes a function floatx80_mod that
is extremely similar to floatx80_rem, but computing the remainder
based on truncating the quotient toward zero rather than rounding it
to nearest integer. This is also useful for emulating the x87 fprem
and fprem1 instructions. Change the floatx80_rem implementation into
floatx80_modrem that can perform either operation, with both
floatx80_rem and floatx80_mod as thin wrappers available for all
targets.
There does not appear to be any use for the _mod operation for other
floating-point formats in QEMU (the only other architectures using
_rem at all are linux-user/arm/nwfpe, for FPA emulation, and openrisc,
for instructions that have been removed in the latest version of the
architecture), so no change is made to the code for other formats.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <alpine.DEB.2.21.2006081654280.23637@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
|
|
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>
|
|
This includes *_is_any_nan, *_is_neg, *_is_inf, etc.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Replace the floatx80 compare specializations with inline functions
that call the standard floatx80_compare{,_quiet} functions.
Use bool as the return type.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Replace the float128 compare specializations with inline functions
that call the standard float128_compare{,_quiet} functions.
Use bool as the return type.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Replace the float64 compare specializations with inline functions
that call the standard float64_compare{,_quiet} functions.
Use bool as the return type.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Replace the float32 compare specializations with inline functions
that call the standard float32_compare{,_quiet} functions.
Use bool as the return type.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Give the previously unnamed enum a typedef name. Use it in the
prototypes of compare functions. Use it to hold the results
of the compare functions.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Give the previously unnamed enum a typedef name. Use the packed
attribute so that we do not affect the layout of the float_status
struct. Use it in the prototypes of relevant functions.
Adjust switch statements as necessary to avoid compiler warnings.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Slightly tidies the usage within softfloat.c and the
representation in float_status.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
We have had this on the to-do list for quite some time.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
The existing f{32,64}_addsub_post test, which checks for zero
inputs, is identical to f{32,64}_mul_fast_test. Which means
we can eliminate the fast_test/fast_op hooks in favor of
reusing the same post hook.
This means we have one fewer test along the fast path for multiply.
Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
The softfloat function floatx80_round_to_int incorrectly handles the
case of a pseudo-denormal where only the high bit of the significand
is set, ignoring that bit (treating the number as an exact zero)
rather than treating the number as an alternative representation of
+/- 2^-16382 (which may round to +/- 1 depending on the rounding mode)
as hardware does. Fix this check (simplifying the code in the
process).
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Message-Id: <alpine.DEB.2.21.2005042339420.22972@digraph.polyomino.org.uk>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
The softfloat floatx80 comparisons fail to allow for pseudo-denormals,
which should compare equal to corresponding values with biased
exponent 1 rather than 0. Add an adjustment for that case when
comparing numbers with the same sign.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Message-Id: <alpine.DEB.2.21.2005042338470.22972@digraph.polyomino.org.uk>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
The softfloat function addFloatx80Sigs, used for addition of values
with the same sign and subtraction of values with opposite sign, fails
to handle the case where the two values both have biased exponent zero
and there is a carry resulting from adding the significands, which can
occur if one or both values are pseudo-denormals (biased exponent
zero, explicit integer bit 1). Add a check for that case, so making
the results match those seen on x86 hardware for pseudo-denormals.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Message-Id: <alpine.DEB.2.21.2005042337570.22972@digraph.polyomino.org.uk>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Conversions between IEEE floating-point formats should convert
signaling NaNs to quiet NaNs. Most of those in QEMU's softfloat code
do so, but those for floatx80 fail to. Fix those conversions to
silence signaling NaNs as well.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Message-Id: <alpine.DEB.2.21.2005042336170.22972@digraph.polyomino.org.uk>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
All other calls to normalize*Subnormal detect zero input before
the call -- this is the only outlier. This case can happen with
+0.0 + +0.0 = +0.0 or -0.0 + -0.0 = -0.0, so return a zero of
the correct sign.
Reported-by: Coverity (CID 1421991)
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Message-Id: <20200327232042.10008-1-richard.henderson@linaro.org>
Message-Id: <20200403191150.863-8-alex.bennee@linaro.org>
|
|
Reintroduce float32_to_float64 that was removed here:
https://lists.gnu.org/archive/html/qemu-devel/2018-04/msg00455.html
- nbench test it not actually calling this function at all
- SPECS 2006 significat number of tests impoved their runtime, just
few of them showed small slowdown
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Matus Kysel <mkysel@tachyum.com>
Message-Id: <20191017142133.59439-1-mkysel@tachyum.com>
[rth: Add comment about impossible inexact exceptions.]
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
This is not a normal header and should only be included in the main
softfloat.c file to bring in the various target specific
specialisations. Indeed as it contains non-inlined C functions it is
not even a legal header. Rename it to match our included C convention.
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
|
|
In our quest to eliminate the home rolled LIT64 macro we fixup usage
inside the softfloat code. While we are at it we remove some of the
extraneous spaces to closer fit the house style.
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Remove some more use of LIT64 while making the meaning more clear. We
also avoid the need of casts as the results by definition fit into the
return type.
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
|
|
This also allows us to remove the extractFloat16exp/frac helpers. We
avoid using the floatXX_pack_raw functions as they are slight overkill
for masking out all but the top bit of the number. The generated code
is almost exactly the same as makes no difference to the
pre-conversion code.
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
We have a wrapper that does the right thing from stdint.h so lets use
it for our constants in softfloat-specialize.h
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Before falling back to softfloat FMA, we do not restore the original
values of inputs A and C. Fix it.
This bug was caught by running gcc's testsuite on RISC-V qemu.
Note that this change gives a small perf increase for fp-bench:
Host: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
Command: perf stat -r 3 taskset -c 0 ./fp-bench -o mulAdd -p $prec
- $prec = single:
- before:
101.71 MFlops
102.18 MFlops
100.96 MFlops
- after:
103.63 MFlops
103.05 MFlops
102.96 MFlops
- $prec = double:
- before:
173.10 MFlops
173.93 MFlops
172.11 MFlops
- after:
178.49 MFlops
178.88 MFlops
178.66 MFlops
Signed-off-by: Kito Cheng <kito.cheng@gmail.com>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Message-Id: <20190322204320.17777-1-cota@braap.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
Wrong type of NaN was generated for IEEE 754-2008 by MADDF.<D|S> and
MSUBF.<D|S> instructions when the arguments were (Inf, Zero, NaN) or
(Zero, Inf, NaN).
The if-else statement establishes if the system conforms to IEEE
754-1985 or IEEE 754-2008, and defines different behaviors depending
on that. In case of IEEE 754-2008, in mentioned cases of inputs,
<MADDF|MSUBF>.<D|S> returns the input value 'c' [2] (page 53) and
raises floating point exception 'Invalid Operation' [1] (pages 349,
350).
These scenarios were tested and the results in QEMU emulation match
the results obtained on the machine that has a MIPS64R6 CPU.
[1] MIPS Architecture for Programmers Volume II-a: The MIPS64
Instruction Set Reference Manual, Revision 6.06
[2] MIPS Architecture for Programmers Volume IV-j: The MIPS64
SIMD Architecture Module, Revision 1.12
Signed-off-by: Mateja Marjanovic <mateja.marjanovic@rt-rk.com>
Message-Id: <1553008916-15274-2-git-send-email-mateja.marjanovic@rt-rk.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
[AJB: fixed up commit message]
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
Previously this was only supported for roundAndPackFloat64.
New support in round_canonical, round_to_int, float128_round_to_int,
roundAndPackFloat32, roundAndPackInt32, roundAndPackInt64,
roundAndPackUint64. This does not include any of the floatx80 routines,
as we do not have users for that rounding mode there.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20190215170225.15537-1-richard.henderson@linaro.org>
Tested-by: David Hildenbrand <david@redhat.com>
[AJB: add missing break]
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
Handling it just like float128_to_uint32_round_to_zero, that hopefully
is free of bugs :)
Documentation basically copied from float128_to_uint64
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
The added branch to the FMA ops is marked as unlikely and therefore
its impact on performance (measured with fp-bench) is within noise range
when measured on an Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz.
Reported-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
|
|
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>
|
|
Performance results for fp-bench:
Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
- before:
sqrt-single: 42.30 MFlops
sqrt-double: 22.97 MFlops
- after:
sqrt-single: 311.42 MFlops
sqrt-double: 311.08 MFlops
Here USE_FP makes a huge difference for f64's, with throughput
going from ~200 MFlops to ~300 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>
|
|
Performance results for fp-bench:
1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
- before:
fma-single: 74.73 MFlops
fma-double: 74.54 MFlops
- after:
fma-single: 203.37 MFlops
fma-double: 169.37 MFlops
2. ARM Aarch64 A57 @ 2.4GHz
- before:
fma-single: 23.24 MFlops
fma-double: 23.70 MFlops
- after:
fma-single: 66.14 MFlops
fma-double: 63.10 MFlops
3. IBM POWER8E @ 2.1 GHz
- before:
fma-single: 37.26 MFlops
fma-double: 37.29 MFlops
- after:
fma-single: 48.90 MFlops
fma-double: 59.51 MFlops
Here having 3FP64 set to 1 pays off for x86_64:
[1] 170.15 vs [0] 153.12 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>
|
|
Performance results for fp-bench:
1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
- before:
div-single: 34.84 MFlops
div-double: 34.04 MFlops
- after:
div-single: 275.23 MFlops
div-double: 216.38 MFlops
2. ARM Aarch64 A57 @ 2.4GHz
- before:
div-single: 9.33 MFlops
div-double: 9.30 MFlops
- after:
div-single: 51.55 MFlops
div-double: 15.09 MFlops
3. IBM POWER8E @ 2.1 GHz
- before:
div-single: 25.65 MFlops
div-double: 24.91 MFlops
- after:
div-single: 96.83 MFlops
div-double: 31.01 MFlops
Here setting 2FP64_USE_FP to 1 pays off for x86_64:
[1] 215.97 vs [0] 62.15 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>
|
|
Performance results for fp-bench:
1. Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
- before:
mul-single: 126.91 MFlops
mul-double: 118.28 MFlops
- after:
mul-single: 258.02 MFlops
mul-double: 197.96 MFlops
2. ARM Aarch64 A57 @ 2.4GHz
- before:
mul-single: 37.42 MFlops
mul-double: 38.77 MFlops
- after:
mul-single: 73.41 MFlops
mul-double: 76.93 MFlops
3. IBM POWER8E @ 2.1 GHz
- before:
mul-single: 58.40 MFlops
mul-double: 59.33 MFlops
- after:
mul-single: 60.25 MFlops
mul-double: 94.79 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>
|
|
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>
|
|
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>
|
|
glibc >= 2.25 defines canonicalize in commit eaf5ad0
(Add canonicalize, canonicalizef, canonicalizel., 2016-10-26).
Given that we'll be including <math.h> soon, prepare
for this by prefixing our canonicalize() with sf_ to avoid
clashing with the libc's canonicalize().
Reported-by: Bastian Koppelmann <kbastian@mail.uni-paderborn.de>
Tested-by: Bastian Koppelmann <kbastian@mail.uni-paderborn.de>
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>
|
|
Older versions of Clang (before 3.5) and GCC (before 4.1) do not
support the "__attribute__((flatten))" yet. We don't care about
such old versions of GCC anymore, but since Clang 3.4 is still
used in EPEL for RHEL7 / CentOS 7, we should not use this attribute
directly but with a wrapper macro instead.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Thomas Huth <thuth@redhat.com>
|
|
The __udiv_qrnnd primitive that we nicked from gmp requires its
inputs to be normalized. We were not doing that. Because the
inputs are nearly normalized already, finishing that is trivial.
Replace div128to64 with a "proper" udiv_qrnnd, so that this
remains a reusable primitive.
Fixes: cf07323d494
Fixes: https://bugs.launchpad.net/qemu/+bug/1793119
Tested-by: Emilio G. Cota <cota@braap.org>
Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Our minimum required compiler for compiling QEMU is GCC 4.1 these days,
so we can drop the support for compilers which do not provide the
__builtin_clz*() functions yet. Since the countLeadingZeros32/64 are
then identical to the clz32/64 functions, and we do not have to sync
the softloat 2 codebase with upstream anymore (softloat 3 is a complete
rewrite) we can simply replace the functions with our QEMU versions.
Suggested-by: Peter Maydell <peter.maydell@linaro.org>
Acked-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Signed-off-by: Thomas Huth <thuth@redhat.com>
Message-Id: <1538118095-7003-1-git-send-email-thuth@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
It has not had users since f83311e476 ("target-m68k: use floatx80
internally", 2017-06-21).
Note that no other bit-width has floatX_trunc_to_int.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20180814002653.12828-3-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
|
|
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20180814002653.12828-2-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
|
|
For 0x1.0000000000003p+0 + 0x1.ffffffep+14 = 0x1.0001fffp+15
we dropped the sticky bit and so failed to raise inexact.
Reported-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Tested-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Message-id: 20180810193129.1556-7-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
|
|
Isolate the target-specific choice to 3 functions instead of 6.
The code in floatx80_default_nan tried to be over-general. There are
only two targets that support this format: x86 and m68k. Thus there
is no point in inventing a mechanism for snan_bit_is_one.
Move routines that no longer have ifdefs out of softfloat-specialize.h.
Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|