[Feature]: Need to dump the GPU assembly code generated with Windows HIP SDK

[wdx04]

Suggestion Description

It seems the HIP compiler on Linux can be controlled by an environment variable KMDUMPISA, when setting KMDUMPISA=1, the compiler can generate assembly code of the GPU code object.

I tried setting the same KMDUMPISA environment variable on Windows, with HIP SDK 5.7.1, but no assembly code was generated.

I need to see the assembly code to find some opportunity for optimization of my code.

Thanks.

Operating System

Windows 10

GPU

RX 7900XT

ROCm Component

HIPCC

[wdx04]

Problem solved: when "-save-temps" option is used, the compiler can generate the assembly code.

However, by viewing the assembly code I found another problem: the sqrtf() function call is too heavily guarded with a total number of 15~22 additional instruction before and after the v_sqrt_f32_e32 instruction, depending on the target architecture. The sqrtf() function in my code is used to calculate the distance of two 2D points: sqrt((x1-x2)**2+(y1-y2)**2). I think it's safe to remove these additional instructions.