Add mkl::linalg_svd relative Ops #1511
Conversation
yucai-intel
commented
Mar 26, 2025
•
yucai-intel
commented
- linalg_svd
- linalg_svd.U
|
Test cases related to complex data types fail because onednn does not support it. |
yucai-intel
changed the title
There was a problem hiding this comment.
Please reactivate some skipped cases such as https://github.com/intel/torch-xpu-ops/blob/main/test/xpu/skip_list_common.py#L1397. Otherwise, CI cannot test with these svd cases.
| #if defined(USE_ONEMKL) | ||
| #include <ATen/native/BatchLinearAlgebra.h> | ||
| #include <ATen/native/DispatchStub.h> | ||
| #include <ATen/native/Resize.h> | ||
| #include <ATen/native/TensorIterator.h> | ||
| #include <ATen/native/xpu/mkl/BatchLinearAlgebra.h> | ||
|
|
||
| #include <ATen/ops/_linalg_svd.h> | ||
| #include <ATen/ops/_linalg_svd_meta.h> | ||
| #include <ATen/ops/_linalg_svd_native.h> | ||
|
|
||
| namespace at::native { | ||
|
|
||
| REGISTER_XPU_DISPATCH(svd_stub, &native::xpu::svd_mkl); | ||
|
|
||
| } // namespace at::native | ||
| #endif // USE_ONEMKL |
There was a problem hiding this comment.
When USE_ONEMKL=OFF, how can we call linalg_svd on XPU?
There was a problem hiding this comment.
@yucai-intel _linalg_svd.U is registered directly instead of using stub. All svd related changes has been removed from native_functions.yaml.
There was a problem hiding this comment.
For ops registered to stub, we will keep their fallback in XPUFallback.template for USE_ONEMKL=OFF.
There was a problem hiding this comment.
Added float64 related test cases.
Local tests are all passed or automatically skipped.
CuiYifeng
force-pushed
the
slogdet
branch
3 times, most recently
from
2a0518c to
56cebb3
Compare
