-
1.16.5 protected0dd64568 · ·
- ``dft`` add benchmark with default hyper-parameters. - Improve performance of ``fuse``. - Improve ``bitrev`` performance. - Add new function ``dft_ks_values``. - Change the API of Chain class. - Fix bug in ``_time_ksmm`` when params is ``None``. - Fix bug in fp16 kernel. - Fix CUDA context stack not empty. - Split ``dft`` function into default ``dft`` and ``dft_helper``. - Review of ``wip.butterfly`` documentation.
-
1.16.4 protected8e10697b · ·
- Add kernel for bit-reversal permutation. - Add ``ks_values` attribute to LazyLinOp returned by ``ksm``. - ``ksd`` does not return list of factors anymore. - Add ``dft``, ``dft_square_dyadic_ks_values`` and ``fuse`` functions to ``wip.butterfly`` module.
-
1.16.3 protected7f763e98 · ·
- Merge ``multiple_ksm()`` into ``ksm()`` function. - Review of the documentation of the ``wip.butterfly`` module.
-
-
-
1.16.1 protected8e3dc48b · ·
- Hyper-parameters TX and TY are multiple of VSIZE now. - wip.butterlfy.ksd: read and save handle complex numbers now. - Module wip.butterlfy: fix a bug when batch size is one.
-
1.16.0 protected2f140456 · ·
- Better management of pyCUDA context and device pointers. - Add conditions on hyper-parameters and block/grid sizes. - Fix a bug in CUDA kernel when dtype is complex.
-
1.15.0 protectede7773261 · ·
- Kronecker product $L_1\otimes L_2$ uses as a *definition* the mixed Kronecker matrix-vector property $(L_2^T\otimes L_1)\mathtt{colvec}(X)=\mathtt{colvec}(L_1XL_2)$. - Introduces ``colvec`` to stack columns of a 2d array. -
-
-
-
-
-
-
-
-
-
-
-