This commit includes hip-7.0.2-fix-libcxx-noinline.patch, which solves
an old issue with libc++. At some point of time this problem should be
reported to upstream, but at this point it is unclear who should fix it
(llvm-project or ROCm), so there is no report. This patch allows to use
rocm_use_clang function from rocm.eclass instead of rocm_use_hipcc.
ROCm gradually deprecates hipcc for simplification and due to
side-effects. Migration to clang was done just recently in the
https://github.com/ROCm/TheRock repository; now Gentoo uses clang as
well (mostly).
Bug: https://bugs.gentoo.org/964799
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44165
Signed-off-by: Sam James <sam@gentoo.org>
This is a small package which includes header and verson library.
Previously the same header was added by dev-util/hip ebuild,
but other packages like amdsmi also use rocm-core library, so it is
better to install them in a separate packages (as intended by upstream).
Bug: https://bugs.gentoo.org/964799
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44165
Signed-off-by: Sam James <sam@gentoo.org>
In ROCm-7 dev-util/rocm-smi reaches maintenance mode (only critical bug fixes will be applied).
A new package dev-util/amdsmi replaces it (includes amd-smi cli tool, libraries and python bindings) and new features and hardware are added only there.
Bug: https://bugs.gentoo.org/964799
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44165
Signed-off-by: Sam James <sam@gentoo.org>
This is initial change for ROCm-7.0.2 packages, followed by registration
of new targets in rocm.eclass.
Most of requests related to gfx115x (Strix / Strix Point) can be
solved only with ROCm-7.0.x (which was added to profiles/desc earlier,
but caused compilation failures). Examples:
https://bugs.gentoo.org/962820,
https://bugs.gentoo.org/962822
This release also adds support for gfx1200 (RDNA4) with few mid-segment
GPUs (https://en.wikipedia.org/wiki/Radeon_RX_9000_series), with <=16GB
of memory, but with native support of int8/int4 quantized computations
(RDNA4-only), which compensates on that.
The new gfx950 target - a flagship in AMD Instinct series (which
requires >64 of RAM to compile, see https://bugs.gentoo.org/960513,
so I can't even compile it on my main machine).
As for gfx1150, the code is still arriving to develop branches.
gfx940 and gfx941 were moved to "unofficial", but these devices should
be compatible with gfx942 code (as far as I know).
While ROCm-7 release was advertised as "Up to 3.8x Performance Uplift",
in reality performance improvements for consumer GPUs are negligible.
Maybe the comparison was done between old fp16 vs new quantized
int8/int4 (i. e. apples with oranges), which is obviously faster.
There are no LLVM changes (still custom version of llvm-20), the release
looks mostly incremental (apart from int8/int4 features, added only for
RDNA4-capable GPUs). There is an internal restructuration, most of the
packages changed HOMEPAGE: hipBLAS, hipBLASLt , hipCUB, hipFFT, hipRAND,
hipSPARSE, hipSPARSELt, MIOpen, rocBLAS, rocFFT, rocPRIM, rocRAND,
rocSPARSE, rocThrust, and Tensile were moved to
https://github.com/ROCm/rocm-libraries monorepository, while some others
were moved to https://github.com/ROCm/rocm-systems. SRC_URI are still
the same, as ROCm continues to publish them. Downloading the whole
monorepo (471 mb tar.gz) is just bad user experience, so we might at
least defer it.
Bug: https://bugs.gentoo.org/964799
Bug: https://bugs.gentoo.org/962820
Bug: https://bugs.gentoo.org/962822
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44165
Signed-off-by: Sam James <sam@gentoo.org>
Sv. Lockal
Michał Górny