[Libre-soc-dev] qemu cosimulation

Jacob Lifshay programmerjake at gmail.com
Wed Jun 2 18:06:16 BST 2021


On Wed, Jun 2, 2021, 05:38 Luke Kenneth Casson Leighton <lkcl at lkcl.net>
wrote:

> first actual discrepancy found: FP operations have to take place
> at the 64 bit precision... then be truncated to IEEE754 FP32,
> then be *re-expanded* to fit into IEEE754 FP64.
>

no, actually f*s (f32) instructions convert all input regs to f32, then do
the operation in f32 format, then the results are converted back to f64.
This *does* give different results than what you guessed above for some fp
operations, such as fma.

Jacob


More information about the Libre-soc-dev mailing list