bae9a58f5d
Even though the regular softmax shaders successfully pass test-backend-ops with Apple GPUs, running long inference tests has shown the models end derailing with softmax OPs being the root cause. With this commit, we use simpler softmax shaders borrowed from the Kompute backend (which are basically reimplementations of the Metal shaders) on certain GPUs know to have problem with the regular ones. Signed-off-by: Sergio Lopez <slp@redhat.com> |
||
|---|---|---|
| .. | ||
| cmake | ||
| include | ||
| src | ||
| .gitignore | ||
| CMakeLists.txt | ||