0.4.2

test the linux wheeel against cpu-only pytorch

0.4.1

export PyInit__C with PyMODINIT_FUNC for Windows

0.4.0

This release comes with new kernel improvements, better benchmark and tests, and some bug fixes in edge cases:

• Various improvements to the CPU and CUDA kernels (6b5d6d1, 9bf3a92, 15da1c4, 4f22bc8)
• Add DeviceGuard to work with multiple GPUs (84b73c6)
• Fix cudaStream (600a2c8)
• Fix rounding errors in accumulation with half-precision (15da1c4)
• Major overhaul of the benchmark (d7efc7e, cdff359)
• Add lots of tests (694e96b, 829ffda, 6f9e497, 67da5e7)
• Auto-generate API reference in README (1569b2d)

0.3.0

Faster CUDA kernel and dispatch to other dtypes

• Add a script to benchmark kernels between torchdtw version: benchmark/src/dtw_benchmark/compare_release.py
• Dispatch to all kinds of dtypes for distances and sx in CPU and CUDA kernels.
• Faster CUDA kernel:

[---------------- dtw / cpu -----------------]
                 |  released 0.2.0  |  current
1 threads: -----------------------------------
      64x64      |         9.2      |      9.4
      128x128    |        23.9      |     24.4
      256x256    |        85.6      |     86.1
      512x512    |       327.1      |    327.8
      1024x1024  |      1291.0      |   1295.6

Times are in microseconds (us).

[-------------------- dtw_batch / cpu --------------------]
                              |  released 0.2.0  |  current
1 threads: ------------------------------------------------
      n=4x4, s=128x128, sym   |        72.8      |     72.8
      n=4x4, s=128x128, asym  |       155.6      |    154.9
      n=8x8, s=256x256, sym   |       665.9      |    663.5
      n=4x8, s=256x512, asym  |      1899.2      |   1901.0

Times are in microseconds (us).

[---------------- dtw / cuda ----------------]
                 |  released 0.2.0  |  current
1 threads: -----------------------------------
      64x64      |        96.7      |     81.3
      128x128    |       149.4      |    123.6
      256x256    |       288.0      |    228.2
      512x512    |       745.2      |    560.8
      1024x1024  |      2581.7      |   1159.4

Times are in microseconds (us).

[-------------------- dtw_batch / cuda -------------------]
                              |  released 0.2.0  |  current
1 threads: ------------------------------------------------
      n=4x4, s=128x128, sym   |      100.0       |    81.3
      n=4x4, s=128x128, asym  |      104.2       |    84.1
      n=8x8, s=256x256, sym   |      305.3       |   153.3
      n=4x8, s=256x512, asym  |      514.8       |   353.0

0.2.0: revert CUDA version in CI, and install 12.9 instead of 13.0

PyTorch version downloaded by `uv build` seems to be the minimum one

0.1.1

no build isolation to get CUDA support on Windows

0.1.0

update tests

0.0.2

fix GA permissions

0.0.1

up