summaryrefslogtreecommitdiffstats
path: root/src/gallium/auxiliary/util/u_linear.h
diff options
context:
space:
mode:
authorRoland Scheidegger <[email protected]>2012-09-21 17:03:48 +0200
committerJosé Fonseca <[email protected]>2012-10-12 18:51:18 +0100
commitd366520e8553f4a16151ee946d6e8136cab3de5e (patch)
treec2f1df1a2710f2c1e4157a40a9a9623d85912bfa /src/gallium/auxiliary/util/u_linear.h
parent2a4105cbc067dcd38057a877d606e9493e9ed53a (diff)
gallivm: fix rsqrt failures
lp_build_rsqrt initially did not do any newton-raphson step. This meant that precision was only ~11 bits, but this handled both input 0.0 and +infinity correctly. It did not however handle input 1.0 accurately, and denormals always generated infinity result. Doing a newton-raphson step increased precision significantly (but notably input 1.0 still doesn't give output 1.0), however this fails for inputs 0.0 and infinity (both result in NaNs). Try to fix this up by using cmp/select but since this is all quite fishy (and still doesn't handle denormals) disable for now. Note that even with workarounds it should still have been faster since the fallback uses sqrt/div (which both use the usually unpipelined and slow divider hw). Also add some more test values to lp_test_arit and test lp_build_rcp() too while there. v2: based on José's feedback, avoid hacky infinity definition which doesn't work with msvc (unfortunately using INFINITY won't cut it neither on non-c99 compilers) in lp_build_rsqrt, and while here fix up the input infinity case too (it's disabled anyway). Only test infinity input case if we have c99, and use float cast for calculating reference rsqrt value so we really get what we expect. Reviewed-by: José Fonseca <[email protected]>
Diffstat (limited to 'src/gallium/auxiliary/util/u_linear.h')
0 files changed, 0 insertions, 0 deletions