v0.2.3
版本发布时间: 2022-06-16 08:43:02
NVIDIA/warp最新发布版本:v1.3.3(2024-09-05 04:54:55)
[0.2.3] - 2022-06-13
- Fix for incorrect 4d array bounds checking
- Fix for
wp.constant
changes not updating module hash - Fix for stale CUDA kernel cache when CPU kernels launched first
- Array gradients are now allocated along with the arrays and accessible as
wp.array.grad
, users should take care to always callwp.Tape.zero()
to clear gradients between different invocations ofwp.Tape.backward()
- Added
wp.array.fill_()
to set all entries to a scalar value (4-byte values only currently)
Breaking Changes
- Tape
capture
option has been removed, users can now capture tapes inside existing CUDA graphs (e.g.: inside Torch) - Scalar loss arrays should now explicitly set
requires_grad=True
at creation time
[0.2.2] - 2022-05-30
- Fix for
from import *
inside Warp initialization - Fix for body space velocity when using deforming Mesh objects with scale
- Fix for noise gradient discontinuities affecting
wp.curlnoise()
- Fix for
wp.from_torch()
to correctly preserve shape - Fix for URDF parser incorrectly passing density to scale parameter
- Optimizations for startup time from 3s -> 0.3s
- Add support for custom kernel cache location, Warp will now store generated binaries in the user's application directory
- Add support for cross-module function references, e.g.: call another modules @wp.func functions
- Add support for overloading
@wp.func
functions based on argument type - Add support for calling built-in functions directly from Python interpreter outside kernels (experimental)
- Add support for auto-complete and docstring lookup for builtins in IDEs like VSCode, PyCharm, etc
- Add support for doing partial array copys, see
wp.copy()
for details - Add support for accessing mesh data directly in kernels, see
wp.mesh_get_point()
,wp.mesh_get_index()
,wp.mesh_eval_face_normal()
- Change to only compile for targets where kernel is launched (e.g.: will not compile CPU unless explicitly requested)
Breaking Changes
- Builtin methods such as
wp.quat_identity()
now call the Warp native implementation directly and will return awp.quat
object instead of NumPy array - NumPy implementations of many builtin methods have been moved to
warp.utils
and will be deprecated - Local
@wp.func
functions should not be namespaced when called, e.g.: previouslywp.myfunc()
would work even ifmyfunc()
was not a builtin - Removed
wp.rpy2quat()
, please usewp.quat_rpy()
instead
[0.2.1] - 2022-05-11
- Fix for unit tests in Kit
1、 warp-public-binary-0.2.3+master.444.44438a8f.teamcity.zip 86MB