[Issue]: Build failure in TensileCreateExtOpLibraries

[LunNova]

Problem Description

I'm trying to package hipblaslt for NixOS and having an issue with the TensileCreateExtOpLibraries("${PROJECT_BINARY_DIR}/Tensile/library" "${Tensile_ARCHITECTURE}") part of the build.

There are a large number of errors at this step for multiple arches:

/build/source/build/library/build_tmp/ops/L_256_4_0_gfx908.s:825:70: error: 'null' operand is not supported on this GPU
buffer_load_dword v[vgprValue], v[vgprOffset], s[sgprSrc:sgprSrc+3], null offen offset:0

/build/source/build/library/build_tmp/ops/A_S_S_256_4_gfx90a.s:204:1: error: operands are not valid for this GPU or mode
v_add_u32 v[vgprOffset], vcc, v[vgprOffset], s[sgprTmp]
^

/build/source/build/library/build_tmp/ops/L_256_4_1_gfx942.s:174:1: error: operands are not valid for this GPU or mode
v_add_u32 v[vgprOffset+3], vcc, v[vgprOffset+2], s[sgprTmp]


/build/source/build/library/build_tmp/ops/L_256_4_0_gfx1100.s:813:26: error: invalid operand for instruction
v_add_u32 v[vgprOffset], vcc, v[vgprOffset], s[sgprTmp]
                         ^

Full log on rocm-6.3.1: https://gist.github.com/LunNova/9534a8a1edd66b23dbe721356abb8c88
Full log on develop 2025-01-19: https://gist.github.com/LunNova/96c368674f531dc576b1995b9087e9b6

I can get the build to succeed by commenting out the call to TensileCreateExtOpLibraries in library/src/amd_detail/rocblaslt/src/CMakeLists.txt, but obviously that's not a correct solution.

This issue is occurring both on the rocm-6.3.1 tag and on the latest develop commit (01839fe) as of 2025-01-19. Building with ROCm 6.3.1 compiler and CLR etc.

Operating System

NixOS 25.05 (Warbler)

CPU

AMD Ryzen 9 7950X 16-Core Processor, EPYC MILAN

GPU

AMD Instinct MI210, AMD Instinct MI100

ROCm Version

ROCm 6.3.1

[lucbruni-amd]

Hi @LunNova,

Just wanted to let you know that an internal ticket has been created and we are investigating. Thanks!

[lucbruni-amd]

Hi @LunNova,

Thanks for providing your build logs. I wasn't able to reproduce this with a fresh install of ROCm 6.3.1 (nor 6.3.2).

Could you kindly provide additional reproduction steps, any environment variables, etc.?

Thank you, and I appreciate your patience.