search

nvdla / sw

NVDLA SW
Other
489 stars 193 forks source link

nvdla_compiler option: per-kernel, per-filter #148

Open zklc opened 5 years ago

zklc commented 5 years ago

"./nvdla_compiler -h" prints "--quantizationMode <per-kernel|per-filter>", so what is the difference between them? Thank you.

zklc commented 5 years ago

Also a question, what does "--profile <basic|default|performance|fast-math> " mean? Since nvdla_compiler does not show the public source code yet, I have to ask you the offical for answer. Thank you!

zklc commented 5 years ago

@jwise the officer, Thank you

prasshantg commented 5 years ago

@zklc per-kernel means we calculate one quantization scale for full weights while per-filter means scale is calculated per-filter.

NVDLA compiler does multiple optimizations and profile indicates level of optimization.