Is this error-rate in sse1 chirp-functions on the new sah 7.00 Mac version expected ? (Tested on Mac mini (mid 2011 model))

setiathome_v7 7.00 XCode GCC 4.0.1 (Apple Inc. build 5494) i386

libboinc: 7.0.58
libboinc: 7.0.58

Work Unit Info:
...............
WU true angle range is : 0.775000
Optimal function choices:
--------------------------------------------------------
name timing error
--------------------------------------------------------
v_BaseLineSmooth (no other)

v_GetPowerSpectrum 0.000226 0.00000 test
v_vGetPowerSpectrum 0.000097 0.00000 test
v_vGetPowerSpectrum2 0.000090 0.00000 test
v_vGetPowerSpectrumUnrolled 0.000091 0.00000 test
v_vGetPowerSpectrumUnrolled2 0.000077 0.00000 test
v_vGetPowerSpectrumUnrolled2 0.000077 0.00000 choice

v_ChirpData 0.007092 0.00000 test
fpu_ChirpData 0.016489 0.00000 test
fpu_opt_ChirpData 0.006907 0.00000 test
v_vChirpData_x86_64 0.070698 0.03947 test
sse1_ChirpData_ak 0.009096 1.51106 test
sse1_ChirpData_ak8e 0.008000 1.51106 test
sse1_ChirpData_ak8h 0.008222 1.51106 test
sse2_ChirpData_ak 0.009883 0.00000 test
sse2_ChirpData_ak8 0.005512 0.00000 test
sse3_ChirpData_ak 0.010804 0.00000 test
sse3_ChirpData_ak8 0.005302 0.00000 test
sse3_ChirpData_ak8 0.005302 0.00000 choice

v_Transpose 0.008579 0.00000 test
v_Transpose2 0.008421 0.00000 test
v_Transpose4 0.004602 0.00000 test
v_Transpose8 0.007673 0.00000 test
v_pfTranspose2 0.011674 0.00000 test
v_pfTranspose4 0.004819 0.00000 test
v_pfTranspose8 0.008436 0.00000 test
v_vTranspose4 0.003427 0.00000 test
v_vTranspose4np 0.003832 0.00000 test
v_vTranspose4ntw 0.007219 0.00000 test
v_vTranspose4x8ntw 0.003296 0.00000 test
v_vTranspose4x16ntw 0.001491 0.00000 test
v_vpfTranspose8x4ntw 0.007619 0.00000 test
v_vTranspose4x16ntw 0.001491 0.00000 choice

FPU opt folding 0.001879 0.00000 test
ben SSE folding 0.001277 0.00000 test
AK SSE folding 0.000993 0.00000 test
BH SSE folding 0.001062 0.00000 test
AK SSE folding 0.000993 0.00000 choice

Test duration 4.43 seconds

---

_\|/_
U r s

I'm sure it's related to the angle reduction. SSE1 doesn't have double precision capability needed there so it's done with x87 code, but the quickround is tricky since Linux operates the FPU at extended precision while Windows and at least some BSD variants use double precision. For OSX it may be using a last resort fallthrough code.

When I was developing that code I ignored Mac-Intel because all those systems have at least SSE2, and that appears true even for OSx86 Hackintosh systems. Eric did some further modifications and cleanups.

The testing is doing what it should, eliminating function variants which are either inaccurate or slow. At least there's no need for immediate code changes.

                                                                   Joe

ok, thanks for explaining that bit. Will keep looking for other unexpected outcomes.

---

_\|/_
U r s