ROCm / MIOpen

AMD's Machine Intelligence Library
https://rocm.docs.amd.com/projects/MIOpen/en/latest/
Other
1.08k stars 230 forks source link

Elaborate experimental FP16 mode of MPBidirectWinograd #521

Open atamazov opened 4 years ago

atamazov commented 4 years ago

Each configuration should be controlled separately. Group convolution needs to be double checked. Then 16/16 F(2,3) and F(3,3) to be tested on a network level. I guess those could be enabled for normal use.

Originally posted by @Kirpich30000 in https://github.com/ROCmSoftwarePlatform/MIOpen/pull/491#issuecomment-710587689

ppanchad-amd commented 8 months ago

@atamazov Is this ticket still relevant?

atamazov commented 8 months ago

@ppanchad-amd Yes, the effect of enabing MIOPEN_DEBUG_AMD_MP_BD_WINOGRAD_EXPEREMENTAL_FP16_TRANSFORM envvar it not investigated yet, and this is quite amount of work to do.