The powerpc finite optimization do not show much gain:
- GCC will call libm iff -fsignaling-nans is used. This usage pattern
is usually not performance oriented and for such calls PLT overhead
should dominate execution time.
- The power7 uses ftdiv to optimize for some input patterns, but at
cost of others. Comparing against generic C implementation built
for powerpc64-linux-gnu-power7 (--with-cpu=power7):
- Generic sysdeps/ieee754 implementation:
"isfinite": {
"": {
"duration": 5.0082e+09,
"iterations": 2.45299e+09,
"max": 43.824,
"min": 2.008,
"mean": 2.04167
},
"INF": {
"duration": 4.66554e+09,
"iterations": 2.28288e+09,
"max": 35.73,
"min": 2.008,
"mean": 2.04371
},
"NAN": {
"duration": 4.66274e+09,
"iterations": 2.28716e+09,
"max": 34.161,
"min": 2.009,
"mean": 2.03866
}
}
- power7 optimized one:
"isfinite": {
"": {
"duration": 4.99111e+09,
"iterations": 2.65566e+09,
"max": 25.015,
"min": 1.716,
"mean": 1.87942
},
"INF": {
"duration": 4.6783e+09,
"iterations": 2.0999e+09,
"max": 35.264,
"min": 1.868,
"mean": 2.22787
},
"NAN": {
"duration": 4.67915e+09,
"iterations": 2.08678e+09,
"max": 38.099,
"min": 1.869,
"mean": 2.24228
}
}
So it basically optimizes marginally for normal numbers while
increasing the latency for other kind of FP.
- The power8 implementation is just the generic implementation using
ISA 2.07 mfvsrd instruction (which GCC uses for generic implementation).
So generic implementation is the best option for powerpc64le.
Checked on powerpc-linux-gnu (built without --with-cpu, with
--with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch),
powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+
and --disable-multi-arch).
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/Makefile
(sysdeps_routines, libm-sysdep_routines): Remove s_finite*
objects.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_finite-power7.S:
Remove file.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_finite-ppc32.c:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_finite.c: Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_finitef-ppc32.c:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_finitef.c: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_finite.S: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_finitef.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile (sysdep_call):
Remove s_finite* objects.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finite-power7.S: Remove file.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finite-power8.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finite-ppc64.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finite.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finitef-ppc64.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_finitef.c: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_finite.S: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_finitef.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_finite.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_finitef.S: Likewise.
Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>
The powerpc isinf optimizations onyl adds complexity:
- GCC will call libm iff -fsignaling-nans is used. This usage pattern
is usually not performance oriented and for such calls PLT overhead
should dominate execution time.
- The power7 uses ftdiv to optimize for some input pattern and branch
implementation for INF and denormal that does:
return (ix & UINT64_C (0x7fffffffffffffff)) == UINT64_C (0x7ff0000000000000)
Although it does show slight better latency than generic algorithm
(as below), it is only for power7 and requires it to override it
for power8.
- The power8 implementation is just the generic implementation using
ISA 2.07 mfvsrd instruction (which GCC uses for generic implementation).
So generic implementation is the best option for powerpc64le.
Checked on powerpc-linux-gnu (built without --with-cpu, with
--with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch),
powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+
and --disable-multi-arch).
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/Makefile
(sysdeps_routines, libm-sysdep_routines): Remove s_isinf* and s_isinf*
objects.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isinf-power7.S:
Remove file.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isinf-ppc32.c:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isinf.c: Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isinff-ppc32.c:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isinff.c: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_isinf.S: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_isinff.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile (sysdep_call):
Remove s_isinf* and s_isinf* objects.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinf-power7.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinf-power8.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinf-ppc64.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinf.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinff-ppc64.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isinff.c: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_isinf.S: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_isinff.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_isinf.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_isinff.S: Likewise.
Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>
The powerpc isnan optimizations are not really a gain:
- GCC will call libm iff -fsignaling-nans is used. This usage pattern
is usually not performance oriented and for such calls PLT overhead
should dominate execution time.
- The power5, power6, and power6x are just micro-optimization to
improve the Load-Hit-Store hazards from floating-point to general
register transfer, and current GCC already has support to minimize
it by inserting either extra nops or group dispatch instructions.
- The power7 uses ftdiv to optimize for some input patterns, but at
cost of others. Comparing against generic C implementation built
for powerpc-linux-gnu-power4 (which uses the hp-timing support on
benchtests):
- Generic sysdeps/ieee754 implementation:
"isnan": {
"": {
"duration": 4.98415e+09,
"iterations": 2.34516e+09,
"max": 45.925,
"min": 2.052,
"mean": 2.12529
},
"INF": {
"duration": 4.74057e+09,
"iterations": 1.69761e+09,
"max": 91.01,
"min": 2.052,
"mean": 2.79249
},
"NAN": {
"duration": 4.74071e+09,
"iterations": 1.68768e+09,
"max": 282.343,
"min": 2.052,
"mean": 2.809
}
}
- power7 optimized one:
$ ./testrun.sh benchtests/bench-isnan
"isnan": {
"": {
"duration": 4.96842e+09,
"iterations": 2.56297e+09,
"max": 50.048,
"min": 1.872,
"mean": 1.93854
},
"INF": {
"duration": 4.76648e+09,
"iterations": 1.54213e+09,
"max": 373.408,
"min": 2.661,
"mean": 3.09084
},
"NAN": {
"duration": 4.76845e+09,
"iterations": 1.54515e+09,
"max": 51.016,
"min": 2.736,
"mean": 3.08607
}
}
So it basically optimizes marginally for normal numbers while
increasing the latency for other kind of FP.
- The generic implementation requires getting the floating point
status, disable the invalid operation bit, and restore the
floating-point status. Each operation is costly and requires
flushing the FP pipeline.
Using the same scenarion for the previous analysis:
"isnan": {
"": {
"duration": 5.08284e+09,
"iterations": 6.2898e+08,
"max": 41.844,
"min": 8.057,
"mean": 8.08108
},
"INF": {
"duration": 4.97904e+09,
"iterations": 6.16176e+08,
"max": 39.661,
"min": 8.057,
"mean": 8.08055
},
"NAN": {
"duration": 4.98695e+09,
"iterations": 5.95866e+08,
"max": 29.728,
"min": 8.345,
"mean": 8.36925
}
}
- The power8 implementation is just the generic implementation using
ISA 2.07 mfvsrd instruction (which GCC uses for generic implementation).
So generic implementation is the best option for powerpc64le.
Checked on powerpc-linux-gnu (built without --with-cpu, with
--with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch),
powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+
and --disable-multi-arch).
* sysdeps/powerpc/fpu/s_isnan.c: Remove file.
* sysdeps/powerpc/fpu/s_isnanf.S: Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/Makefile
(sysdeps_routines, libm-sysdep_routines): Remove s_isnan-* and
s_isnanf-* objects.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnan-power5.S:
Remove file
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnan-power6.S:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnan-power7.S:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnan-ppc32.S:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnan.c: Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnanf-power5.S:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnanf-power6.S:
Likewise.
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_isnanf.c: Likewise.
* sysdeps/powerpc/powerpc32/power5/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc32/power5/fpu/s_isnanf.S: Likewise.
* sysdeps/powerpc/powerpc32/power6/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc32/power6/fpu/s_isnanf.S: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc32/power7/fpu/s_isnanf.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile (sysdep_calls):
Remove s_isnan-* and s_isnanf-* objects.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-power5.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-power6.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-power6x.S:
Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-power7.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-power8.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan-ppc64.S: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnan.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/s_isnanf.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power5/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power6/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power6x/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power7/fpu/s_isnanf.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_isnan.S: Likewise.
* sysdeps/powerpc/powerpc64/power8/fpu/s_isnanf.S: Likewise.
Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>
This patch provides optimized logb (1.2x on PPC32 and 2.5x on PPC64),
logbf (1.1x on PPC32 and 2.2x on PPC64), and logbl (1.3x on PPC32 and
50% on PPC64) for the POWER7 processor.