sci-libs/rocWMMA
library for accelerating mixed precision matrix multiply-accumulate operations
Screenshots
-
rocWMMA-10.1.0~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 +amdgpu_targets_gfx1250 test
View 
Download 
Browse License: MIT Overlay: stuff -
rocWMMA-10.1~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: inode64-overlay -
rocWMMA-10.0.0~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: inode64-overlay -
rocWMMA-10.0.0~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 +amdgpu_targets_gfx1250 test
View Download Browse License: MIT Overlay: stuff -
rocWMMA-7.2.4~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: inode64-overlay -
rocWMMA-7.2.0~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1152 +amdgpu_targets_gfx1153 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: gentoo -
rocWMMA-7.1.0~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: gentoo -
rocWMMA-7.0.2~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: gentoo -
rocWMMA-6.4.3~amd64+amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx940 +amdgpu_targets_gfx941 +amdgpu_targets_gfx942 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 test
View Download Browse License: MIT Overlay: gentoo
ChangeLog
commit c30c3f17005457d379fa6c200be05a8b4d388334
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Oct 9 23:06:24 2026 +0200
sci-libs/rocWMMA: add 10.1.0
10.1 probes the compiler for --offload-jobs or -parallel-jobs itself
and no longer passes -fallow-half-arguments-and-returns, so the
src_prepare that stripped both flags has nothing left to do.
commit 77997eefb90129c6e03cb3f93bb91ec718458471
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:39:58 2026 +0200
sci-libs/rocWMMA: trim 10.0.0 comment prose
commit 4cebe971e14c00bfeed6139fc204cb59062fe489
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:48:30 2026 +0200
sci-libs/rocWMMA: new package, add 10.0.0
Last ROCm component ::gentoo carried that this overlay did not. ::gentoo stops
at 7.2.0, and rocWMMA pins dev-util/hip:$, so its 7.2.0 cannot solve
against the 10.0 stack -- leaving the host with no rocWMMA at all. Left
unmasked for the same reason dev-util/amdsmi and dev-libs/rocdbgapi are: the
10.0 version is the only one that can solve here, so masking it would remove
the package rather than defer it.
Forked from ::gentoo's 7.2.0 and retargeted to the rocm-libraries
therock-<major.minor> release asset, since AMD retired the rocm-* line at
rocm-7.2.4.
Three deltas against the ::gentoo ebuild, all re-read from the therock-10.0
source rather than assumed:
- dev-util/rocm-smi dropped. Upstream migrated to amdsmi at 10.0 --
find_package(amd_smi) in CMakeLists.txt, <amd_smi/amdsmi.h> in
test/hip_device.cpp -- and no rocm_smi reference survives anywhere in the
tree. It also moved out of the library's own dependencies: the
find_package sits behind ROCWMMA_BUILD_BENCHMARK_TESTS, and with that off
AMD_SMI_LIBRARY stays empty so the INTERFACE target links nothing extra.
Only the test binaries need it, via hip_device.cpp in
ROCWMMA_COMMON_TEST_SOURCES, so it is test-gated here.
- gfx1250 added, matching upstream's DEFAULT_GPU_TARGETS list. gfx1103 is
deliberately NOT added: it has a ROCWMMA_ARCH_GFX1103 macro in
internal/config.hpp but no entry in the CMake target list, so it is a
source-level arch predicate rather than a shipped build target.
- gtest floored at 1.16.0, which is what test/CMakeLists.txt now requires.
Both hipcc-flag seds gained grep-or-die guards. sed exits 0 on no-match, so
an upstream flag removal would leave them silently inert and surface later as
an unrelated-looking hipcc error.
Build-verified and run-verified on gfx1150: a 16x16x16 f16 WMMA GEMM compiled
with hipcc against the installed headers returns the exact expected result in
all 256 output elements. Confirmed after src_prepare that both seds and the
forked no-test-install patch actually took effect, and that nothing lands in
/usr/bin.
commit a5947a70afb079cfb4f60bef381f6beabf08798e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Dec 6 21:57:18 2025 +0100
*/*: drop rocm/hip as newer version is in main tree, cntd
Signed-off-by: Ivan S. Titov <iohann.s.titov@gmail.com>
commit cfa997ed467a25f0a98e30e69523013fb12b4136
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Oct 18 14:26:09 2025 +0200
added rocm and HIP v. 6.4.4 for now, as main tree lags
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Oct 9 23:06:24 2026 +0200
sci-libs/rocWMMA: add 10.1.0
10.1 probes the compiler for --offload-jobs or -parallel-jobs itself
and no longer passes -fallow-half-arguments-and-returns, so the
src_prepare that stripped both flags has nothing left to do.
commit 77997eefb90129c6e03cb3f93bb91ec718458471
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:39:58 2026 +0200
sci-libs/rocWMMA: trim 10.0.0 comment prose
commit 4cebe971e14c00bfeed6139fc204cb59062fe489
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:48:30 2026 +0200
sci-libs/rocWMMA: new package, add 10.0.0
Last ROCm component ::gentoo carried that this overlay did not. ::gentoo stops
at 7.2.0, and rocWMMA pins dev-util/hip:$, so its 7.2.0 cannot solve
against the 10.0 stack -- leaving the host with no rocWMMA at all. Left
unmasked for the same reason dev-util/amdsmi and dev-libs/rocdbgapi are: the
10.0 version is the only one that can solve here, so masking it would remove
the package rather than defer it.
Forked from ::gentoo's 7.2.0 and retargeted to the rocm-libraries
therock-<major.minor> release asset, since AMD retired the rocm-* line at
rocm-7.2.4.
Three deltas against the ::gentoo ebuild, all re-read from the therock-10.0
source rather than assumed:
- dev-util/rocm-smi dropped. Upstream migrated to amdsmi at 10.0 --
find_package(amd_smi) in CMakeLists.txt, <amd_smi/amdsmi.h> in
test/hip_device.cpp -- and no rocm_smi reference survives anywhere in the
tree. It also moved out of the library's own dependencies: the
find_package sits behind ROCWMMA_BUILD_BENCHMARK_TESTS, and with that off
AMD_SMI_LIBRARY stays empty so the INTERFACE target links nothing extra.
Only the test binaries need it, via hip_device.cpp in
ROCWMMA_COMMON_TEST_SOURCES, so it is test-gated here.
- gfx1250 added, matching upstream's DEFAULT_GPU_TARGETS list. gfx1103 is
deliberately NOT added: it has a ROCWMMA_ARCH_GFX1103 macro in
internal/config.hpp but no entry in the CMake target list, so it is a
source-level arch predicate rather than a shipped build target.
- gtest floored at 1.16.0, which is what test/CMakeLists.txt now requires.
Both hipcc-flag seds gained grep-or-die guards. sed exits 0 on no-match, so
an upstream flag removal would leave them silently inert and surface later as
an unrelated-looking hipcc error.
Build-verified and run-verified on gfx1150: a 16x16x16 f16 WMMA GEMM compiled
with hipcc against the installed headers returns the exact expected result in
all 256 output elements. Confirmed after src_prepare that both seds and the
forked no-test-install patch actually took effect, and that nothing lands in
/usr/bin.
commit a5947a70afb079cfb4f60bef381f6beabf08798e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Dec 6 21:57:18 2025 +0100
*/*: drop rocm/hip as newer version is in main tree, cntd
Signed-off-by: Ivan S. Titov <iohann.s.titov@gmail.com>
commit cfa997ed467a25f0a98e30e69523013fb12b4136
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Oct 18 14:26:09 2025 +0200
added rocm and HIP v. 6.4.4 for now, as main tree lags