About division on amdgpu backend

[0oyyo0]

I'm looking for information on the AMD GPU backend's support for integer division. I haven't found any instructions related to integer division in the ISA documentation, and the code I see for division, such as in AMDGPUCodeGenPrepare.cpp and SIISelLowering.cpp, uses other methods for calculation, e.g. (a/b---->a*(1/b)) . Therefore, I'd like to understand the AMD GPU hardware's support for division and why it's converted to other methods. I suspect it's because division is time-consuming.

One conversion method is as follows:

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Helium (0oyyo0)

I'm looking for information on the AMD GPU backend's support for integer division. I haven't found any instructions related to integer division in the ISA documentation, and the code I see for division, such as in AMDGPUCodeGenPrepare.cpp and SIISelLowering.cpp, uses other methods for calculation, e.g. (a/b---->a*(1/b)) . Therefore, I'd like to understand the AMD GPU hardware's support for division and why it's converted to other methods. I suspect it's because division is time-consuming.

One conversion method is as follows:
<img width="1138" height="1061" alt="Image" src="https://github.com/user-attachments/assets/cc106d72-af97-475d-997b-acb3151667db" />

[arsenm]

There is no hardware support. It requires an expansion

[0oyyo0]

There is no hardware support. It requires an expansion

What about floating point support? I can find instructions related to floating point division in the ISA documentation.

[arsenm]

What about floating point support? I can find instructions related to floating point division in the ISA documentation.

There's also no direct floating point division instruction. v_div_fmas_* and v_div_scale_* are intended to help with this expansion, but aren't a direct division instruction.

[0oyyo0]

What about floating point support? I can find instructions related to floating point division in the ISA documentation.

There's also no direct floating point division instruction. v_div_fmas_* and v_div_scale_* are intended to help with this expansion, but aren't a direct division instruction.

OK, thanks for your reply.

[shiltian]

It seems like the question has been answered. Are we good to close this out now?