Changelog

[1.6.2] - 2025-03-07

Changed

• Update project license fromNVIDIA Software License toApache License, Version 2.0 (seeLICENSE.md).

Changelog

[1.6.1] - 2025-03-03

Added

• Documentwp.Launch objects (docs,#428).
• Document how overwriting previously computed results can lead to incorrect gradients (docs,#525).

Fixed

• Fix unaligned loads with offset 2D tiles inwp.tile_load().
• Fix FP64 accuracy of thread-level matrix-matrix multiplications (#489).
• Fixwp.array() not initializing from arrays defining a CUDA array interface when the target device is CPU (#523).
• Fixwp.Launch objects not storing and replaying adjoint kernel launches (#449).
• Fixwp.config.verify_autograd_array_access failing to detect overwrites in generic Warp functions (#493).
• Fix an error on Windows when closing anOpenGLRenderer app (#488).
• Fix per-vertex colors not being correctly written out to USD meshes when a constant color is being passed (#480).
• Fix an error in capturing thewp.sim.VBDIntegrator with CUDA graphs whenhandle_self_contact is enabled (#441).
• Fix an error of AABB computation inwp.collide.TriMeshCollisionDetector.
• Fix URDF-imported planar joints not being set with the intendedtarget_ke,target_kd, andmode parameters (#454).
• FixModelBuilder.add_builder() to use correct offsets forModelBuilder.joint_parent andModelBuilder.joint_child (#432)
• Fix underallocation of contact points for box–sphere and box–capsule collisions.
• Fixwp.randi() documentation to show correct output range of[-2^31, 2^31).

Changelog

[1.6.0] - 2025-02-03

Added

• Add preview of Tile Cholesky factorization and solve APIs throughwp.tile_cholesky(),tile_cholesky_solve()
  andtile_diag_add() (preview APIs are subject to change).
• Support for loading tiles from arrays whose shapes are not multiples of the tile dimensions.
  Out-of-bounds reads will be zero-filled and out-of-bounds writes will be skipped.
• Support for higher-dimensional (up to 4D) tile shapes and memory operations.
• Add intersection-free self-contact support inwp.sim.VDBIntegrator by passinghandle_self_contact=True.
  Seewarp/examples/sim/example_cloth_self_contact.py for a usage example.
• Add functionswp.norm_l1(),wp.norm_l2(),wp.norm_huber(),wp.norm_pseudo_huber(), andwp.smooth_normalize()
  for vector types to a newwp.math module.
• wp.sim.SemiImplicitIntegrator andwp.sim.FeatherstoneIntegrator now have an optionalfriction_smoothing
  constructor argument (defaults to 1.0) that controls softness of the friction norm computation.
• Supportassert statements in kernels (docs).
  Assertions can only be triggered in"debug" mode (GH-366).
• Support CUDA IPC on Linux. Call theipc_handle() method to get an IPC handle for awp.Event or awp.array,
  and callwp.from_ipc_handle() orwp.event_from_ipc_handle() in another process to open the handle
  (docs).
• Add per-module option to disable fused floating point operations, usewp.set_module_options({"fuse_fp": False})
  (GH-379).
• Add per-module option to add CUDA-C line information for profiling, usewp.set_module_options({"lineinfo": True}).
• Support operator overloading forwp.struct objects by definingwp.func functions
  (GH-392).
• Add built-in functionwp.len() to retrieve the number of elements for vectors, quaternions, matrices, and arrays
  (GH-389).
• Addwarp/examples/optim/example_softbody_properties.py as an optimization example for soft-body properties
  (GH-419).
• Addwarp/examples/tile/example_tile_walker.py, which reworks the existingexample_walker.py
  to use Warp's tile API for matrix multiplication.
• Addwarp/examples/tile/example_tile_nbody.py as an example of an N-body simulation using Warp tile primitives.

Changed

• Breaking: Changewp.tile_load() andwp.tile_store() indexing behavior so that indices are now specified in
  terms ofarray elements instead oftile multiples.
• Breaking: Tile operations now takeshape andoffset parameters as tuples,
  e.g.:wp.tile_load(array, shape=(m,n), offset=(i,j)).
• Breaking: Change exception types and error messages thrown by tile functions for improved consistency.
• Add an implicit tile synchronization whenever a shared memory tile's data is reinitialized (e.g. in dynamic loops).
  This could result in lower performance.
• wp.Bvh constructor now supports various construction algorithms via theconstructor argument, including
  "sah" (Surface Area Heuristics),"median", and"lbvh" (docs)
• Improve the query efficiency ofwp.Bvh andwp.Mesh.
• Improve memory consumption, compilation and runtime performance when using in-place vector/matrix assignments in
  kernels that haveenable_backward set toFalse (GH-332).
• Vector/matrix/quaternion component+= and-= operations compile and run faster in the backward pass
  (GH-332).
• Name files in the kernel cache according to their directory. Previously, all files began with
  module_codegen (GH-431).
• Avoid recompilation of modules when changingblock_dim.
• wp.autograd.gradcheck_tape() now has additional optional argumentsreverse_launches andskip_to_launch_index.
• wp.autograd.gradcheck(),wp.autograd.jacobian(), andwp.autograd.jacobian_fd() now also accept
  arbitrary Python functions that have Warp arrays as inputs and outputs.
• update_vbo_transforms kernel launches in the OpenGL renderer are no longer recorded onto the tape.
• Skip emitting backward functions/kernels in the generated C++/CUDA code whenenable_backward is set toFalse.
• Emit deprecation warnings for the use of theowner andlength keywords in thewp.array initializer.
• Emit deprecation warnings for the use ofwp.mlp(),wp.matmul(), andwp.batched_matmul().
  Use tile primitives instead.

Fixed

• Fix unintended modification of non-Warp arrays during the backward pass (GH-394).
• Fix so thatwp.Tape.zero() zeroes gradients passed via thegrads parameter inwp.Tape.backward()
  (GH-407).
• Fix errors during graph capture caused by module unloading (GH-401).
• Fix potential memory corruption errors when allocating arrays with strides (GH-404).
• Fixwp.array() not respecting the targetdtype andshape when the given data is an another array with a CUDA interface
  (GH-363).
• Negative constants evaluate to compile-time constants (GH-403)
• FixImportError exception being thrown during interpreter shutdown on Windows when using the OpenGL renderer
  (GH-412).
• Fix the OpenGL renderer not working when multiple instances exist at the same time (GH-385).
• FixAttributeError crash in the OpenGL renderer when moving the camera (GH-426).
• Fix the OpenGL renderer not correctly displaying duplicate capsule, cone, and cylinder shapes
  (GH-388).
• Fix the overriding ofwp.sim.ModelBuilder default parameters (GH-429).
• Fix indexing ofwp.tile_extract() when the block dimension is smaller than the tile size.
• Fix scale and rotation issues with the rock geometry used in the granular collision SDF example
  (GH-409).
• Fix autodiff Jacobian computation inwp.autograd.jacobian() where in some cases gradients were not zeroed-out properly.
• Fix plotting issues inwp.autograd.jacobian_plot().
• Fix thelen() operator returning the total size of a matrix instead of its first dimension.
• Fix gradient instability in rigid-body contact handling forwp.sim.SemiImplicitIntegrator and
  wp.sim.FeatherstoneIntegrator (GH-349).
• Fix overload resolution of generic Warp functions with default arguments.
• Fix rendering of arrows with differentup_axis,color inOpenGLRenderer (GH-448).

Changelog

[1.5.1] - 2025-01-02

Added

• Add PyTorch basics and custom operators notebooks to thenotebooks directory.
• Update PyTorch interop docs to include section on custom operators
  (docs).

Fixed

• warp.sim: Fix a bug in which the color-balancing algorithm was not updating the colorings.
• Fix custom colors being not being updated when rendering meshes with static topology in OpenGL
  (GH-343).
• Fixwp.launch_tiled() not returning aLaunch object when passedrecord_cmd=True.
• Fix default arguments not being resolved forwp.func when called from Python's runtime
  (GH-386).
• Array overwrite tracking: Fix issue with not marking arrays passed towp.atomic_add(),wp.atomic_sub(),
  wp.atomic_max(), orwp.atomic_min() as being written to (GH-378).
• Fix for occasional failure to update.meta files into Warp kernel cache on Windows.
• Fix the OpenGL renderer not being able to run without a CUDA device available
  (GH-344).
• Fix incorrect CUDA driver function versions (GH-402).

Changelog

[1.5.0] - 2024-12-02

Added

• Support for cooperative tile-based primitives using cuBLASDx and cuFFTDx, please see the tile
  documentation for details.
• Expose areversed() built-in for iterators (GH-311).
• Support for saving Volumes into.nvdb files with thesave_to_nvdb method.
• warp.fem: Addwp.fem.Trimesh3D andwp.fem.Quadmesh3D geometry types for 3D surfaces with newexample_distortion_energy example.
• warp.fem: Add"add" option towp.fem.integrate() for accumulating integration result to existing output.
• warp.fem: Add"assembly" option towp.fem.integrate() for selecting between more memory-efficient or more
  computationally efficient integration algorithms.
• warp.fem: Add Nédélec (first kind) and Raviart-Thomas vector-valued function spaces
  providing conforming discretization ofcurl anddiv operators, respectively.
• warp.sim: Add a graph coloring module that supports converting trimesh into a vertex graph and applying coloring.
  Thewp.sim.ModelBuilder now includes methods to color particles for use withwp.sim.VBDIntegrator(),
  users should callbuilder.color() before finalizing assets.
• warp.sim: Add support for a per-particle radius for soft-body triangle contact using thewp.sim.Model.particle_radius
  array (docs), replacing the previous
  hard-coded value of 0.01 (GH-329).
• Add aparticle_radius parameter towp.sim.ModelBuilder.add_cloth_mesh() andwp.sim.ModelBuilder.add_cloth_grid()
  to set a uniform radius for the added particles.
• Documentwp.array attributes (GH-364).
• Document time-to-compile tradeoffs when using vector component assignment statements in kernels.
• Add introductory Jupyter notebooks to thenotebooks directory.

Changed

• Drop support for Python 3.7; Python 3.8 is now the minimum-supported version.
• Promote thewp.Int,wp.Float, andwp.Scalar generic annotation types to the public API.
• warp.fem: Simplify querying neighboring cell quantities when integrating on sides using new
  wp.fem.cells(),wp.fem.to_inner_cell(),wp.fem.to_outer_cell() operators.
• Show an error message when the type returned by a function differs from its annotation, which would have led to the compilation stage failing.
• Clarify thatwp.randn() samples a normal distribution of mean 0 and variance 1.
• Raise error when passing more than 32 variadic argument to thewp.printf() built-in.

Fixed

• Fixplace setting of paddle backend.
• warp.fem: Fix tri-cubic shape functions on quadrilateral meshes.
• warp.fem: Fix caching of integrand kernels when changing code-generation options.
• Fixwp.expect_neq() overloads missing for scalar types.
• Fix an error when awp.kernel or awp.func object is annotated to return aNone value.
• Fix error when reading multi-volume, BLOSC-compressed.nvdb files.
• Fixwp.printf() erroring out when no variadic arguments are passed (GH-333).
• Fix memory access issues in soft-rigid contact collisions (GH-362).
• Fix gradient propagation for in-place addition/subtraction operations on custom vector-type arrays.
• Fix the OpenGL renderer's window not closing when clicking the X button.
• Fix the OpenGL renderer's camera snapping to a different direction from the initial camera's orientation when first looking around.
• Fix custom colors being ignored when rendering meshes in OpenGL (GH-343).
• Fix topology updates not being supported by the the OpenGL renderer.

Changelog

[1.4.2] - 2024-11-13

Changed

• Make the output ofwp.print() in backward kernels consistent for all supported data types.

Fixed

• Fix to relax the integer types expected when indexing arrays (regression in1.3.0).
• Fix printing vector and matrix adjoints in backward kernels.
• Fix kernel compile error when printing structs.
• Fix an incorrect user function being sometimes resolved when multiple overloads are available with array parameters with differentdtype values.
• Fix error being raised when static and dynamic for-loops are written in sequence with the same iteration variable names (GH-331).
• Fix an issue with theTexture Write node, used in the Mandelbrot Omniverse sample, sometimes erroring out in multi-GPU environments.
• Code generation of in-place multiplication and division operations (regression introduced ina69d061)(GH-342).

Changelog

[1.4.1] - 2024-10-15

Fixed

• Fixiter_reverse() not working as expected for ranges with steps other than 1 (GH-311).
• Fix potential out-of-bounds memory access when awp.sparse.BsrMatrix object is reused for storing matrices of different shapes.
• Fix robustness to very low desired tolerance inwp.fem.utils.symmetric_eigenvalues_qr.
• Fix invalid code generation error messages when nesting dynamic and static for-loops.
• Fix caching of kernels with static expressions.
• FixModelBuilder.add_builder(builder) to correctly updatearticulation_start and therebyarticulation_count whenbuilder contains more than one articulation.
• Re-introduced thewp.rand*(),wp.sample*(), andwp.poisson() onto the Python scope to revert a breaking change.

CHANGELOG

[1.4.0] - 2024-10-01

Added

• Support for a newwp.static(expr) function that allows arbitrary Python expressions to be evaluated at the time of
  function/kernel definition (docs).
• Support for stream priorities to hint to the device that it should process pending work
  in high-priority streams over pending work in low-priority streams when possible
  (docs).
• Adaptive sparse grid geometry towarp.fem (docs).
• Support for definingwp.kernel andwp.func objects from within closures.
• Support for defining multiple versions of kernels, functions, and structs without manually assigning unique keys.
• Support for default argument values for user functions decorated withwp.func.
• Allow passing custom launch dimensions tojax_kernel() (GH-310).
• JAX interoperability examples for sharding and matrix multiplication (docs).
• Interoperability support for the PaddlePaddle ML framework (GH-318).
• Supportwp.mod() for vector types (GH-282).
• Expose the modulo operator% to Python's runtime scalar and vector types.
• Support for fp64atomic_add,atomic_max, andatomic_min (GH-284).
• Support for quaternion indexing (e.g.q.w).
• Support shadowing builtin functions (GH-308).
• Support for redefining function overloads.
• Add an ocean sample to theomni.warp extension.
• warp.sim.VBDIntegrator now supports body-particle collision.
• Add acontributing guide to the Sphinx docs .
• Add documentation for dynamic code generation (docs).

Changed

• wp.sim.Model.edge_indices now includes boundary edges.
• Unexposedwp.rand*(),wp.sample*(), andwp.poisson() from the Python scope.
• Skip unused functions in module code generation, improving performance.
• Avoid reloading modules if their content does not change, improving performance.
• wp.Mesh.points is now a property instead of a raw data member, its reference can be changed after the mesh is initialized.
• Improve error message when invalid objects are referenced in a Warp kernel.
• if/else/elif statements with constant conditions are resolved at compile time with no branches being inserted in the generated code.
• Include all non-hidden builtins in the stub file.
• Improve accuracy of symmetric eigenvalues routine inwarp.fem.

Fixed

• Fix forwp.func erroring out when defining aTuple as a return type hint (GH-302).
• Fix array in-place op (+=,-=) adjoints to compute gradients correctly in the backwards pass
• Fix vector, matrix in-place assignment adjoints to compute gradients correctly in the backwards pass, e.g.:v[1] = x
• Fix a bug in which Python docstrings would be created as local function variables in generated code.
• Fix a bug with autograd array access validation in functions from different modules.
• Fix a rare crash during error reporting on some systems due to glibc mismatches.
• Handle--num_tiles 1 inexample_render_opengl.py (GH-306).
• Fix the computation of body contact forces inFeatherstoneIntegrator when bodies and particles collide.
• Fix bug inFeatherstoneIntegrator whereeval_rigid_jacobian could give incorrect results or reach an infinite
  loop when the body and joint indices were not in the same order. AddedModel.joint_ancestor to fix the indexing
  from a joint to its parent joint in the articulation.
• Fix wrong vertex index passed toadd_edges() called fromModelBuilder.add_cloth_mesh() (GH-319).
• Add a workaround for uninitialized memory read warning in thecompute-sanitizer initcheck tool when usingwp.Mesh.
• Fix name clashes when Warp functions and structs are returned from Python functions multiple times.
• Fix name clashes between Warp functions and structs defined in different modules.
• Fix code generation errors when overloading generic kernels defined in a Python function.
• Fix issues with unrelated functions being treated as overloads (e.g., closures).
• Fix handling ofstream argument inarray.__dlpack__().
• Fix a bug related to reloading CPU modules.
• Fix a crash when kernel functions are not found in CPU modules.
• Fix conditions not being evaluated as expected inwhile statements.
• Fix printing Boolean and 8-bit integer values.
• Fix array interface type strings used for Boolean and 8-bit integer values.
• Fix initialization error when setting struct members.
• Fix Warp not being initialized upon entering awp.Tape context.
• UsekDLBool instead ofkDLUInt for DLPack interop of Booleans.

[1.3.3] - 2024-09-04

• Bug fixes
  • Fix an aliasing issue with zero-copy array initialization from NumPy introduced in Warp 1.3.0.
  • Fixwp.Volume.load_from_numpy() behavior whenbg_value is a sequence of values.

[1.3.2] - 2024-08-30

• Bug fixes
  • Fix accuracy of 3x3 SVDwp.svd3 with fp64 numbers (GH-281).
  • Fix module hashing when a kernel argument contained a struct array (GH-287).
  • Fix a bug inwp.bvh_query_ray() where the direction instead of the reciprocal direction was used
    (GH-288).
  • Fix errors when launching a CUDA graph after a module is reloaded. Modules that were used during graph capture
    will no longer be unloaded before the graph is released.
  • Fix a bug inwp.sim.collide.triangle_closest_point_barycentric() where the returned barycentric coordinates may be
    incorrect when the closest point lies on an edge.
  • Fix 32-bit overflow when array shape is specified usingnp.int32.
  • Fix handling of integer indices in theinput_output_mask argument toautograd.jacobian and
    autograd.jacobian_fd (GH-289).
  • FixModelBuilder.collapse_fixed_joints() to correctly update the body centers of mass and the
    ModelBuilder.articulation_start array.
  • Fix precedence of closure constants over global constants.
  • Fix quadrature point indexing inwp.fem.ExplicitQuadrature (regression from 1.3.0).
• Documentation improvements
  • Add missing return types for built-in functions.
  • Clarify that atomic operations also return the previous value.
  • Clarify thatwp.bvh_query_aabb() returns parts that overlap the bounding volume.

[1.3.1] - 2024-07-27

• Removewp.synchronize() from PyTorch autograd function example
• Tape.check_kernel_array_access() andTape.reset_array_read_flags() are now private methods.
• Fix reporting unmatched argument types

[1.3.0] - 2024-07-25

• Warp Core improvements

  • Update to CUDA 12.x by default (requires NVIDIA driver 525 or newer), please seeREADME.md for commands to install CUDA 11.x binaries for older drivers
  • Add information to the module load print outs to indicate whether a module was
    compiled(compiled), loaded from the cache(cached), or was unable to be
    loaded(error).
  • wp.config.verbose = True now also prints out a message upon the entry to awp.ScopedTimer.
  • Addwp.clear_kernel_cache() to the public API. This is equivalent towp.build.clear_kernel_cache().
  • Add code-completion support forwp.config variables.
  • Remove usage of a static task (thread) index for CPU kernels to address multithreading concerns (GH-224)
  • Improve error messages for unsupported Python operations such as sequence construction in kernels
  • Updatewp.matmul() CPU fallback to use dtype explicitly innp.matmul() call
  • Add support for PEP 563'sfrom __future__ import annotations (GH-256).
  • Allow passing external arrays/tensors towp.launch() directly via__cuda_array_interface__ and__array_interface__, up to 2.5x faster conversion from PyTorch
  • Add faster Torch interop path usingreturn_ctype argument towp.from_torch()
  • Handle incompatible CUDA driver versions gracefully
  • Addwp.abs() andwp.sign() for vector types
  • Expose scalar arithmetic operators to Python's runtime (e.g.:wp.float16(1.23) * wp.float16(2.34))
  • Add support for creating volumes with anisotropic transforms
  • Allow users to pass function arguments by keyword in a kernel using standard Python calling semantics
  • Add additional documentation and examples demonstratingwp.copy(),wp.clone(), andarray.assign() differentiability
  • Add__new__() methods for all class__del__() methods to handle when a class instance is created but not instantiated before garbage collection
  • Implement the assignment operator forwp.quat
  • Make the geometry-related built-ins available only from within kernels
  • Rename the API-facing query types to remove their_t suffix:wp.BVHQuery,wp.HashGridQuery,wp.MeshQueryAABB,wp.MeshQueryPoint, andwp.MeshQueryRay
  • Addwp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)

• warp.autograd improvements:

  • Newwarp.autograd module with utility functionsgradcheck(),jacobian(), andjacobian_fd() for debugging kernel Jacobians (docs)
  • Add array overwrite detection, ifwp.config.verify_autograd_array_access is true in-place operations on arrays on the Tape that could break gradient computation will be detected (docs)
  • Fix bug where modification of@wp.func_replay functions and native snippets would not trigger module recompilation
  • Add documentation for dynamic loop autograd limitations

• warp.sim improvements:

  • Improve memory usage and performance for rigid body contact handling whenself.rigid_mesh_contact_max is zero (default behavior).
  • Themask argument towp.sim.eval_fk() now accepts both integer and boolean arrays to mask articulations.
  • Fix handling ofModelBuilder.joint_act inModelBuilder.collapse_fixed_joints() (affected floating-base systems)
  • Fix and improve implementation ofModelBuilder.plot_articulation() to visualize the articulation tree of a rigid-body mechanism
  • Fix ShapeInstancer__new__() method (missing instance return and*args parameter)
  • Fix handling ofupaxis variable inModelBuilder and the rendering thereof inOpenGLRenderer

• warp.sparse improvements:

  • Sparse matrix allocations (frombsr_from_triplets(),bsr_axpy(), etc.) can now be captured in CUDA graphs; exact number of non-zeros can be optionally requested asynchronously.
  • bsr_assign() now supports changing block shape (including CSR/BSR conversions)
  • Add Python operator overloads for common sparse matrix operations, e.gA += 0.5 * B,y = x @ C

• warp.fem new features and fixes:

  • Support for variable number of nodes per element
  • Globalwp.fem.lookup() operator now supportswp.fem.Tetmesh andwp.fem.Trimesh2D geometries
  • Simplified defining custom subdomains (wp.fem.Subdomain), free-slip boundary conditions
  • New field types:wp.fem.UniformField,wp.fem.ImplicitField andwp.fem.NonconformingField
  • Newstreamlines,magnetostatics andnonconforming_contact examples, updatedmixed_elasticity to use a nonlinear model
  • Function spaces can now export VTK-compatible cells for visualization
  • Fixed edge cases with NanoVDB function spaces
  • Fixed differentiability ofwp.fem.PicQuadrature w.r.t. positions and measures

[1.3.2] - 2024-08-30

• Bug fixes
  • Fix accuracy of 3x3 SVDwp.svd3 with fp64 numbers (GH-281).
  • Fix module hashing when a kernel argument contained a struct array (GH-287).
  • Fix a bug inwp.bvh_query_ray() where the direction instead of the reciprocal direction was used
    (GH-288).
  • Fix errors when launching a CUDA graph after a module is reloaded. Modules that were used during graph capture
    will no longer be unloaded before the graph is released.
  • Fix a bug inwp.sim.collide.triangle_closest_point_barycentric() where the returned barycentric coordinates may be
    incorrect when the closest point lies on an edge.
  • Fix 32-bit overflow when array shape is specified usingnp.int32.
  • Fix handling of integer indices in theinput_output_mask argument toautograd.jacobian and
    autograd.jacobian_fd (GH-289).
  • FixModelBuilder.collapse_fixed_joints() to correctly update the body centers of mass and the
    ModelBuilder.articulation_start array.
  • Fix precedence of closure constants over global constants.
  • Fix quadrature point indexing inwp.fem.ExplicitQuadrature (regression from 1.3.0).
• Documentation improvements
  • Add missing return types for built-in functions.
  • Clarify that atomic operations also return the previous value.
  • Clarify thatwp.bvh_query_aabb() returns parts that overlap the bounding volume.

[1.3.1] - 2024-07-27

• Removewp.synchronize() from PyTorch autograd function example
• Tape.check_kernel_array_access() andTape.reset_array_read_flags() are now private methods.
• Fix reporting unmatched argument types

[1.3.0] - 2024-07-25

• Warp Core improvements
  • Update to CUDA 12.x by default (requires NVIDIA driver 525 or newer), please seeREADME.md for commands to install CUDA 11.x binaries for older drivers
  • Add information to the module load print outs to indicate whether a module was
    compiled(compiled), loaded from the cache(cached), or was unable to be
    loaded(error).
  • wp.config.verbose = True now also prints out a message upon the entry to awp.ScopedTimer.
  • Addwp.clear_kernel_cache() to the public API. This is equivalent towp.build.clear_kernel_cache().
  • Add code-completion support forwp.config variables.
  • Remove usage of a static task (thread) index for CPU kernels to address multithreading concerns (GH-224)
  • Improve error messages for unsupported Python operations such as sequence construction in kernels
  • Updatewp.matmul() CPU fallback to use dtype explicitly innp.matmul() call
  • Add support for PEP 563'sfrom __future__ import annotations (GH-256).
  • Allow passing external arrays/tensors towp.launch() directly via__cuda_array_interface__ and__array_interface__, up to 2.5x faster conversion from PyTorch
  • Add faster Torch interop path usingreturn_ctype argument towp.from_torch()
  • Handle incompatible CUDA driver versions gracefully
  • Addwp.abs() andwp.sign() for vector types
  • Expose scalar arithmetic operators to Python's runtime (e.g.:wp.float16(1.23) * wp.float16(2.34))
  • Add support for creating volumes with anisotropic transforms
  • Allow users to pass function arguments by keyword in a kernel using standard Python calling semantics
  • Add additional documentation and examples demonstratingwp.copy(),wp.clone(), andarray.assign() differentiability
  • Add__new__() methods for all class__del__() methods to handle when a class instance is created but not instantiated before garbage collection
  • Implement the assignment operator forwp.quat
  • Make the geometry-related built-ins available only from within kernels
  • Rename the API-facing query types to remove their_t suffix:wp.BVHQuery,wp.HashGridQuery,wp.MeshQueryAABB,wp.MeshQueryPoint, andwp.MeshQueryRay
  • Addwp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)