gpu f16 cast to fp32 calculation, and then converted back?
Opened this issue · 1 comments
chengmengli06 commented
For elementwise operations with fp16 input, the data is first converted to fp32, and convert back after call gpu functions? But gpu actually support fp16 and bf16.
bool cast_result_to_fp16 = false;
chengmengli06 commented
besides, does xla support tf32 now?