Apparently ptxas in CUDA 7.0 and 7.5 has a bug which miscompiles
cvt.rni.s64.f32 instruction so lrint(1.f) and llrint(1.f) return 0 instead of 1.
CUDA 8 and in-driver JIT which comes with it do not have this issue.
Paths
| Differential D27741
[test-suite,CUDA] disable [l]lrint(1.f) checks on CUDA 7.0,7.5 ClosedPublic Authored by tra on Dec 13 2016, 4:39 PM.
Details Summary Apparently ptxas in CUDA 7.0 and 7.5 has a bug which miscompiles CUDA 8 and in-driver JIT which comes with it do not have this issue.
Diff Detail
Event Timelinetra updated this object. Closed by commit rL289687: [test-suite,CUDA] disable [l]lrint(1.f) checks on CUDA 7.0,7.5 (authored by tra). · Explain WhyDec 14 2016, 11:06 AM This revision was automatically updated to reflect the committed changes.
Revision Contents
Diff 81424 test-suite/trunk/External/CUDA/cmath.cu
test-suite/trunk/External/CUDA/math_h.cu
|