|
From: Julian S. <js...@ac...> - 2006-04-13 13:36:14
|
> in various stack traces. I've patched filter_libc to fix it. Great. > The interesting failure I noticed (and this isn't a change) is > actually memcheck/tests/x86/sse1_memory which is giving different > output for most of the rcpss/rcpps/rsqrtss/rsqrtps tests on most > of my machines (both 32 and 64 bit). What it is is, these are really reciprocal-estimate/reciprocal-sqrt- estimate instructions. The amd docs say, for rsqrt*s, "the maximum error is <= 1.5 * 2^-12 times the true result", and the same for rcp*s. (saves cycles - if you want more precision you can do another stage of newton-rapson iteration afterwards "by hand"). So the test case is deficient - for those insns the result comparison needs to take into account only the top 11 or 12 bits of the mantissa. J |