[hatari-devel] FPU trig functions emulation accuracy?

[ Thread Index | Date Index | More lists.tuxfamily.org/hatari-devel Archives ]


Hi,

As there's been discussion about FPU, and I found
Douglas' old FPU test from my HD, I gave it
a whirl with current soft FPU.

Results showed:
--------------------------------------
....
fscale:brand    pass
fscale:drand    pass
fscale:erand    pass
fnegabs:brand   pass
fnegabs:drand   pass
fnegabs:erand   pass
ftrig:brand     fail $cc9df7b41129f3cb
ftrig:drand     fail $e60a6aad01e2f19f
ftrig:erand     fail $49917b5e83becf2f
fatrig:brand    fail $a35f4f1f3a15397a
fatrig:drand    fail $33754097f36b27a7
fatrig:erand    fail $187a6208c4c2cd41
ftrigh:brand    fail $a28e04f0b56dcd28
ftrigh:drand    fail $c6f1d06af77f63f4
ftrigh:erand    fail $bcc8427633ed42e8
fytox:brand     fail $6dc3fd64dd855755
fytox:drand     fail $0a82171044f6b1ce
fytox:erand     fail $64d9478527c87218
fexpman:brand   pass
fexpman:drand   pass
fexpman:erand   pass
flog:brand      fail $ce1fe252a81bf58d
flog:drand      fail $36fdff39848684fa
flog:erand      fail $b3be6b3a7fe4d834
fsincos:brand   fail $2e60a9b61bdf61ab
fsincos:drand   fail $6d40411d6c318cb8
fsincos:erand   fail $a4b4eadb845409cd
complete - press any key...
--------------------------------------

Was there ever effort to fix trig errors reported
by Douglas test tool?


	- Eero

PS. Douglas wrote (in 2017) this about the info
output by the test:
-----------------------------------------------
Each test is usually repeated with 3 sources of float data:

BRAND = binary random data, including values which
        could be considered invalid
DRAND = double precision random source values
        from -1.0 -> +1.0
ERAND = same as DRAND, but raised to a random
        exponent between -1024 -> +1023

So errors just in the BRAND tests for example,
could indicate different handling of invalid or
certain denormalised values.

Errors in DRAND tests might indicate general
incorrect handling or general precision
differences in the mantissa.

Errors in ERAND tests might indicate problems with
denormalised values or exponent limits.
-----------------------------------------------



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/