Debian Patches

Status for onednn/3.12.1+ds-3

Patch Description Author Forwarded Bugs Origin Last update
soversion.patch Change SOVERSION to 3.6 due to ukernel API changes Shengqi Chen <harry@debian.org> yes 2025-01-25
system-spdlog.patch Use spdlog from Debian archive Shengqi Chen <harry@debian.org> not-needed 2025-01-24
system-gtest.patch Use googletest from Debian archive Shengqi Chen <harry@debian.org> not-needed 2025-01-24
disable-rvv-riscv64.patch not enable DNNL_RISCV_USE_RVV_INTRINSICS Bo YU <tsu.yubo@gmail.com> not-needed 2024-10-31
disable-benchdnn-tests.patch Disable benchdnn tests which are very slow to compile and run Shengqi Chen <harry@debian.org> not-needed 2025-01-24
enable-loong64-build.patch Add basic build support for loongarch64. Dandan Zhang <zhangdandan@loongson.cn> no debian 2025-04-07
aarch64-cache-per-core-fallback-to-guess.patch aarch64: fall back to guessed cache size when query returns 0 Upstream commit f90453adbf ("cpu: platform: aarch64: query and calculate
cache_per_core") made get_per_core_cache_size() query the real per-core data
cache size instead of a fixed guess. On hosts/environments that do not expose
the L1 *data* cache size (common on VMs, containers and build chroots, and on
some Neoverse parts where /sys lists no L1d size), the query returns 0 (and
the divisor getCoresSharingDataCache() can also be 0).
.
A 0 cache size silently breaks consumers. In particular the aarch64 reorder
planner (prb_block_for_cache) computes L1_cache_sz = 3/4 * 0 = 0, so
"inner_block_sz > L1_cache_sz" is always true and it forces a cache-blocking
transform that yields a kernel which drops a block of data for padded
blocked-weight reorders (the reorder_simple_test_t_f32_f32 / PaddedWeights
failures). simple_concat likewise then always takes its hand-vectorised copy
path (cmp against 0), which overruns its buffer (the int8 concat SIGBUS).
.
Restore the pre-regression behaviour where the size is unavailable: return the
architectural guess() when the queried per-core size is 0, and guard the
divide-by-zero. This fixes both the reorder data-loss and the concat crash on
non-SVE arm64 build environments, with no change where cache size is exposed.
.
Found by git-bisect (v3.9.1 good, v3.12.1 bad) on a Neoverse-N1 host whose
/sys exposes no L1d size; full test_reorder and the int8 concat test pass with
this change.		 Aron Xu <aron@debian.org> yes upstream 2026-06-12

All known versions for source package 'onednn'

Links