Kenneth Heafield
Kenneth Heafield
In any case the writer shouldn't be in multiply.h anymore but that's a longer issue...
@XapaJIaMnu Can you help with your internal compiler error expertise?
Such PRs would be would be welcome. For ARM, we've been using https://github.com/google/ruy though the lack of postprocessing https://github.com/google/ruy/issues/307 is cramping my style (and requiring an extra round-trip to RAM...
Does 65276ad59ab9cd5b2bc623c2411f481f79aa7c5c fix this?
@sidkashyap-at-Intel Can you setup CI for icc?
Sigh there's a bug in the Intel compiler in which constructors with target attributes compile but fail to link. Minimal example: ``` /* Compiles then fails to link */ class...
@XapaJIaMnu says he tried to report this to Intel two years ago via @sidkashyap-at-Intel but couldn't figure out how to raise it to the team.
Target attributes pre-date function multiversioning. Function multiversioning is a feature that should only be activated if there are two function declarations with the same signature that differ only in target...
I tried to find documentation on icc. To quote page 2212 of https://software.intel.com/content/dam/develop/external/us/en/documents/cpp_compiler_classic.pdf : > On Linux*, in addition to the Intel-defined attributes cpu_specific and cpu_dispatch , C++ compilations with...
> When the icc compiler sees a target attribute, it thinks you are using gnu function multiversioning. Since your program didn't provide a "default" function, the multiversion definition wasn't complete....