search

ROCm / ROCm

AMD ROCm™ Software - GitHub Home
https://rocm.docs.amd.com
MIT License
4.56k stars 375 forks source link

gfx1010(RX5000 series) dosen't support fp16 on linux #1857

Open AiPs1717 opened 1 year ago

AiPs1717 commented 1 year ago

currently using gfx1010. In the latest version of ROCm, 5.3.2, gfx1010 cannot use fp16 calculations. An attempt was made to calculate fp16 in the pytorch environment, but it did not work. Is there a way to force the fp16 calculation in the gfx1010 series?

zjin-lcf commented 1 year ago

Could you run the HIP programs without errors in https://github.com/CHIP-SPV/chip-spv/blob/main/samples/fp16 ?

Yougoshatenshi commented 1 year ago

Could you run the HIP programs without errors in https://github.com/CHIP-SPV/chip-spv/blob/main/samples/fp16 ?

Ran all 3 cpp files in hipcc without any errors, but also have the problem that i cant use FP16 in pytorch with ROCm on my 5700XT. In directml on windows its working fine too, although almost as slow as FP32 in pytorch under linux.

abhimeda commented 8 months ago

@AiPs1717 Hi, is this resolved on the latest ROCm? If so can we close this ticket?

Zakhrov commented 5 months ago

@AiPs1717 Hi, is this resolved on the latest ROCm? If so can we close this ticket?

It has been partially, see https://github.com/ROCm/ROCm/issues/2527#issuecomment-2074468176