search

nvdla / sw

NVDLA SW
Other
489 stars 193 forks source link

why nvdla_compiler source code show off a int8 default compute precision value. #209

Closed LeiWang1999 closed 3 years ago

LeiWang1999 commented 3 years ago

I debug for detail of nvdla_compiler, but i found the default presicion of source code is int8, while the nvdia_compiler of nvdla/vp docker images is fp16, i think code of this line is wrong, the default value of cprecision should be nvdla::DataType::HALF

https://github.com/nvdla/sw/blob/79538ba1b52b040a4a4645f630e457fa01839e90/umd/apps/compiler/main.cpp#L62

LeiWang1999 commented 3 years ago

Given my preceding research, this may because that current hw branch was just supported int8 feature.