search

rdaly525 / coreir

BSD 3-Clause "New" or "Revised" License
101 stars 24 forks source link

Why is IEEE compliance for CW floating point add/mul set to 0? #972

Open jack-melchert opened 4 years ago

jack-melchert commented 4 years ago

https://github.com/rdaly525/coreir/blob/367774c87737ad4e09e4cf62c5c7ac6b28825b69/src/libs/float_CW.cpp#L137 https://github.com/rdaly525/coreir/blob/367774c87737ad4e09e4cf62c5c7ac6b28825b69/src/libs/float_CW.cpp#L157

This causes errors when simulating lassen with CW floats: https://github.com/StanfordAHA/lassen/issues/178

Setting ieee_compliance to true fixes this issue. Is there any reason why it was set to false in the first place?

rdaly525 commented 4 years ago

ieee_compliance is essentially a switch indicating compliance with subnormal/denormal floating point values. If you are seeing testing errors, that might mean that we are generating subnormal numbers in the tests. If I recall correctly, we were not trying to support these. Ill try to dig up an email discussing this issue.

rdaly525 commented 4 years ago

That being said, CoreIR should probably expose that flag as a parameter to be decided by the user of the library