Skip to content

Tags: edelsohn/oneDNN

Tags

v1.6.1

Toggle v1.6.1's commit message
fixup: cpu: avx2: zero-pad scratchpad only for nxc format in 1x1 f32 …

…bwd_w driver

v2.0-beta08

Toggle v2.0-beta08's commit message
build: fix an issue with GNU > 9 in TBB cmake

v1.6

Toggle v1.6's commit message
cpu: bnorm: fix for illegal division by 0

v1.6-rc

Toggle v1.6-rc's commit message
fixup: cpu: support ic and oc tails for nxc format in 1x1 f32 avx2 ji…

…t kernel

v1.5.1

Toggle v1.5.1's commit message
benchdnn: add deconv regression test

v2.0-beta07

Toggle v2.0-beta07's commit message
common: primitive cache: add asserts

v1.5

Toggle v1.5's commit message
build: mkldnn compat: put symlinks together with generate lib

CMake allows overriding the place where the libraries are being
generated using:

``` cmake
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY some_place)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY some_place)
```

In this case, the `libdnnl.so` will be put in `some_place` instead of
usual `${CMAKE_CURRENT_BINARY_DIR}` location.

However, we generate the compat symlinks in the latter location. This
makes them point to non-existent library, which in turn:
1. Invalidates the symlink;
2. Betrays users expectations on find the library (which is though
   a symlink) in the proper location `some_place`.
3. Makes `make` regenerate the symlink all the time no matter whether
   it exists or not. That happens because `make` checks whether the
   target (symlink) is _older_ than the prerequisite (`libdnnl.so`).
   It seems that the date for symlink is checked by the date of the
   object it points to. Since the symlink points to nowhere, the
   target is considered always outdated.

This closes oneapi-src#743.

v1.5-rc

Toggle v1.5-rc's commit message
common: nd_iterator: avoid idiv in a loop

v0.21.5

Toggle v0.21.5's commit message
build: bumped version to v0.21.5

v1.4

Toggle v1.4's commit message
fixup: fixup: benchdnn: eltwise: improve input data coverage

* 1) With fp16 argument, exp() is dispatched to fp16 version as well,
* overflows and produces inf instead of an fp32 value, resulting in
* 0 in the final answer, while it should be a non-zero value exceeding trh.
* 2) MSVC math functions are invariant to compiler options leaving the
* answer same no matter what. Current threshold is low for it. Slightly
* increasing it...