ROCm/hipBLASLt

[Issue]: Dependency on Tensile headers

trixirt opened this issue · 3 comments

Problem Description

I am packaging hipBLASLt for Fedora.
When building with -DBUILD_WITH_TENSILE=OFF post 5.7, there are build errors like this

hipBLASLt/library/src/amd_detail/rocblaslt/src/include/tensile_host.hpp:45:10: fatal error: 'Tensile/DataTypes.hpp' file not found
45 | #include <Tensile/DataTypes.hpp>
| ^~~~~~~~~~~~~~~~~~~~~~~

Operating System

Fedora Rawhide

CPU

x86_64

GPU

AMD Instinct MI210

Other

No response

ROCm Version

ROCm 5.7.1

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

the issue tracker drop down on rocm version needs to include 6.0

@trixirt Can you please check if your issue still occurs with the latest ROCm 6.1.2? If not, please close the ticket. Thanks!

@ppanchad-amd , I have the same error the error with 6.1.1/6.1.2, it occurs when user tries to build hipBLASLt with -DBUILD_WITH_TENSILE=OFF (exactly because of this option). The real problem is that this option can not be turned off: not only it fails with fatal error: 'Tensile/DataTypes.hpp' file not found, after adding this header to include path it still fails down the code.

In Gentoo I solved this issue by adding this patch: https://github.com/gentoo/gentoo/blob/master/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-arch.patch. When user has no gfx90a/gfx940/gfx941/gfx942 it allows to build without Tensile as a dummy library just to proceed with pytorch compilation. Otherwise BUILD_WITH_TENSILE must be enabled.

As @AngryLoki says, this still an issue, the option can not be turned off and worse, hipblastlt uses it's own copy of tensile, this conflicts with rocblas's copy.