NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #3187 most downloaded on PyPI
A Python framework for high-performance simulation and graphics programming
Latest stable
1.19.0
Last release 20 days ago
02 Nov 2026
Ships on a steady schedule
a new release about every 4 weeks
Nearly every release is documented
notes for 54 of 54 stable releases
Nothing withdrawn
no release was ever pulled
5 years old
55 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
Remove wp.mlp() (deprecated in v1.6.0). Use tile primitives instead.
wp.map() to map a function over arrays and add math operators for Warp arrays (docs, #694).wp.capture_if() and wp.capture_while() (docs, #597).wp.capture_debug_dot_print() to write a DOT file describing the structure of a captured CUDA graph (#746).Device.sm_count property to get the number of streaming multiprocessors on a CUDA device (#584).wp.block_dim() to query the number of threads in the current block inside a kernel (#695).wp.atomic_cas() and wp.atomic_exch() built-ins for atomic compare-and-swap and exchange operations (#767).wp.config.compile_time_trace setting or the module-level "compile_time_trace" option. When used, JSON files in the Trace Event format will be written in the kernel cache, which can be opened in a viewer like chrome://tracing/ (docs, #609).wp.svd3() and wp.quat_to_axis_angle() (#503).wp.func functions (#682).wp.tile_squeeze() to remove axes of length one (#662).wp.tile_reshape() to reshape a tile (#663).wp.tile_astype() to return a new tile with the same data but different data type. (#683).wp.tile_cholesky_solve() (#773).wp.tile_scan_inclusive() and wp.tile_scan_exclusive() for performing inclusive and exclusive scans over tiles (#731).wp.transform_compose() and wp.transform_decompose() for converting between transforms and 4x4 matrices with 3D scale information (#576).wp.transform syntax operations for loading and storing (#710).as_spheres parameter to UsdRenderer.render_points() in order to choose whether to render the points as USD spheres using a point instancer or as simple USD points (#634).wp.sim.VBDIntegrator.rebuild_bvh() to rebuild the BVH used for detecting self-contacts.wp.sim.VBDIntegrator collisions, with strength is controlled by Model.soft_contact_kd.wp.fem.lookup() operator across geometries and add filtering parameters (#618).warp.fem: fem/example_elastic_shape_optimization.py and fem/example_darcy_ls_optimization.py (#698).py.typed marker file (per PEP 561) to the package to formally support static type checking by downstream users (#780).wp.mlp() (deprecated in v1.6.0). Use tile primitives instead.wp.autograd.plot_kernel_jacobians() (deprecated in v1.4.0). Use wp.autograd.jacobian_plot() instead.length and owner keyword arguments from wp.array() constructor (deprecated in v1.6.0). Use the shape and deleter keywords instead.kernel keyword argument from wp.autograd.jacobian() and wp.autograd.jacobian_fd() (deprecated in v1.6.0). Use the function keyword argument instead.outputs keyword argument from wp.autograd.jacobian_plot() (deprecated in v1.6.0).warp.sim module (planned for removal in v1.10). It will be superseded by the upcoming Newton library, a separate package with a new API. Migrating will require code changes; a future guide will be provided (current draft). See the GitHub announcement for details (#735).wp.matrix(pos, quat, scale) built-in function. Use wp.transform_compose() instead (#576).len() where possible.wp.types.type_length() to wp.types.type_size().wp.tile_cholesky_solve() input parameters to align with its docstring (#726).wp.tile_upper_solve() and wp.tile_lower_solve() to use libmathdx 0.2.1 TRSM solver (#773).wp.tile_matmul() if enable_backward is disabled (#644).preserve_type=True when tiling a value across the block with wp.Tile() (#772).wp.sparse.bsr_[set_]from_triplets differentiable with respect to the input triplet values (#760).warp.fem operators: node_count, node_index, element_coordinates, element_closest_point.wp.sim.VBDIntegrator rigid-body-contact handling to use only the shape's friction coefficient, rather than averaging the shape's and the cloth's coefficients.wp.assign_copy() hidden built-in to the kernel scope.inputs and outputs arguments in the Kernel documentation.wp.launch() by avoiding costly native API calls (#774).@wp.func-decorated functions from the Python scope (#521).Formal parameter space overflowed error during wp.sim.VBDIntegrator kernel compilation for the backward pass in CUDA 11 Warp builds. This was resolved by decoupling collision and elasticity evaluations into separate kernels, increasing parallelism and speeding up the solver (#442).UsdRenderer.render_points() not supporting multiple colors (#634).wp.fem module regarding the orientation of 2D geometry side normals (#629).Your coding agent can read these notes before it upgrades. Set up the MCP server →
Add missing adjoint method for tile assign operations (GH-680).
assign operations (GH-680).+= and -= invoke wp.atomic_add() and wp.atomic_sub(), respectively
(GH-505).wp.struct (now throws RuntimeError)
(GH-656).__cuda_array_interface__ (GH-624,
GH-670).omni.warp towards omni.warp.core
(GH-702).wp.Volume allocation
(GH-611).wp.tile_atomic_add() (GH-681).wp.svd2() with duplicate singular values and improved accuracy
(GH-679).OpenGLRenderer.update_shape_instance() not having color buffers created for the shape instances.wp.render.OpenGLRenderer (GH-704).ModelBuilder.collapse_fixed_joints()
(GH-631).UsdRenderer.render_points() erroring out when passed 4 points or less
(GH-708).wp.atomic_*() built-ins not working with some types (GH-733).Improve handling of deprecated JAX features (#613).
mpi4py in warp/examples/distributed/example_jacobi_mpi.py (#475).repr() for Warp types, including adding repr() for wp.array.framesPerSecond for time sampling instead of timeCodesPerSecond to avoid playback speed issues in some viewers (#617).Model.rigid_contact_tids are now -1 at non-active contact indices which allows to retrieve the vertex index of a mesh collision, see test_collision.py (#623).DeformedGeometry from wp.fem.Trimesh3D geometries (#614).lookup operator for wp.fem.Trimesh3D (#618).dtype parameter missing for wp.quaternion().dtype comparison when using the wp.matrix()/wp.vector()/wp.quaternion() constructors with literal values and an explicit dtype argument (#651).wp.sim.collide() (#459).wp.sim.ModelBuilder adds springs with -1 as vertex indices (#621).show_joints not working with wp.sim.render.SimRenderer set to render to USD (#510).OgnParticlesFromMesh node not being computed correctly.atol and rtol arguments to wp.autograd.gradcheck() and wp.autograd.gradcheck_tape() (#508).Deprecate constructing a matrix from vectors using wp.matrix().
#line directives in CUDA-C code. This setting is controlled by wp.config.line_directives and is True by default. (docs, #437)vec4f grid construction in wp.Volume.allocate_by_tiles().wp.svd2() (#436).wp.randu() for random uint32 generation.wp.matrix_from_cols() and wp.matrix_from_rows() (#278).wp.transform_from_matrix() to obtain a transform from a 4x4 matrix (#211).wp.where() to select between two arguments conditionally using a more intuitive argument order (cond, value_if_true, value_if_false) (#469).wp.get_mempool_used_mem_current() and wp.get_mempool_used_mem_high() to query the respective current and high-water mark memory pool allocator usage (#446 ).Stream.is_complete and Event.is_complete properties to query completion status (#435).wp.clear_lto_cache() to clear the LTO cache (#507).warp/examples/optim/example_fluid_checkpoint.py.wp.sim.VBDIntegrator.wp.matmul() functionality (including batched version). Users should use tile primitives for matrix multiplication operations instead.wp.matrix().wp.select() in favor of wp.where(). Users should update their code to use wp.where(cond, value_if_true, value_if_false) instead of wp.select(cond, value_if_false, value_if_true).wp.sim.Control no longer has a model attribute (#487).wp.sim.Control.reset() is deprecated and now only zeros-out the controls (previously restored controls to initial model state). Use wp.sim.Control.clear() instead.v[0] = x) now compile and run faster in the backward pass. Note: For correct gradient computation, each component should only be assigned once.@wp.kernel has now an optional module argument that allows passing a wp.context.Module to the kernel, or, if set to "unique" let Warp create a new unique module just for this kernel. The default behavior to use the current module is unchanged.wp.tile_reduce() on tiles with struct data types.wp.tile_broadcast() to support broadcasting to 1D, 3D, and 4D shapes (in addition to existing 2D support).wp.fem.integrate() and wp.fem.interpolate() may now perform parallel evaluation of quadrature points within elements.wp.fem.interpolate() can now build Jacobian sparse matrices of interpolated functions with respect to a trial field.wp.sparse routines (bsr_set_from_triplets, bsr_assign, bsr_axpy, bsr_mm) now accept a masked flag to discard any non-zero not already present in the destination matrix.wp.sparse.bsr_assign() no longer requires source and destination block shapes to evenly divide each other.wp.expect_near() to support all vectors and quaternions.wp.quat_from_matrix() to support 4x4 matrices.OgnClothSimulate node to use the VBD integrator (#512).globalScale parameter from the OgnClothSimulate node.edge_indices when adding a ModelBuilder to another (#557).Update project license from *NVIDIA Software License* to *Apache License, Version 2.0* (see LICENSE.md).
Document wp.Launch objects (docs, #428).
wp.Launch objects (docs, #428).wp.tile_load().wp.array() not initializing from arrays defining a CUDA array interface when the target device is CPU (#523).wp.Launch objects not storing and replaying adjoint kernel launches (#449).wp.config.verify_autograd_array_access failing to detect overwrites in generic Warp functions (#493).OpenGLRenderer app (#488).wp.sim.VBDIntegrator with CUDA graphs when handle_self_contact is enabled (#441).wp.collide.TriMeshCollisionDetector.target_ke, target_kd, and mode parameters (#454).ModelBuilder.add_builder() to use correct offsets for ModelBuilder.joint_parent and ModelBuilder.joint_child (#432)wp.randi() documentation to show correct output range of [-2^31, 2^31).Emit deprecation warnings for the use of the owner and length keywords in the wp.array initializer.
wp.tile_cholesky(), tile_cholesky_solve()
and tile_diag_add() (preview APIs are subject to change).wp.sim.VDBIntegrator by passing handle_self_contact=True.
See warp/examples/sim/example_cloth_self_contact.py for a usage example.wp.norm_l1(), wp.norm_l2(), wp.norm_huber(), wp.norm_pseudo_huber(), and wp.smooth_normalize()
for vector types to a new wp.math module.wp.sim.SemiImplicitIntegrator and wp.sim.FeatherstoneIntegrator now have an optional friction_smoothing
constructor argument (defaults to 1.0) that controls softness of the friction norm computation.assert statements in kernels (docs).
Assertions can only be triggered in "debug" mode (GH-366).ipc_handle() method to get an IPC handle for a wp.Event or a wp.array,
and call wp.from_ipc_handle() or wp.event_from_ipc_handle() in another process to open the handle
(docs).wp.set_module_options({"fuse_fp": False})
(GH-379).wp.set_module_options({"lineinfo": True}).wp.struct objects by defining wp.func functions
(GH-392).wp.len() to retrieve the number of elements for vectors, quaternions, matrices, and arrays
(GH-389).warp/examples/optim/example_softbody_properties.py as an optimization example for soft-body properties
(GH-419).warp/examples/tile/example_tile_walker.py, which reworks the existing example_walker.py
to use Warp's tile API for matrix multiplication.warp/examples/tile/example_tile_nbody.py as an example of an N-body simulation using Warp tile primitives.wp.tile_load() and wp.tile_store() indexing behavior so that indices are now specified in
terms of array elements instead of tile multiples.shape and offset parameters as tuples,
e.g.: wp.tile_load(array, shape=(m,n), offset=(i,j)).wp.Bvh constructor now supports various construction algorithms via the constructor argument, including
"sah" (Surface Area Heuristics), "median", and "lbvh" (docs)wp.Bvh and wp.Mesh.enable_backward set to False (GH-332).+= and -= operations compile and run faster in the backward pass
(GH-332).module_codegen (GH-431).block_dim.wp.autograd.gradcheck_tape() now has additional optional arguments reverse_launches and skip_to_launch_index.wp.autograd.gradcheck(), wp.autograd.jacobian(), and wp.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.enable_backward is set to False.owner and length keywords in the wp.array initializer.wp.mlp(), wp.matmul(), and wp.batched_matmul().
Use tile primitives instead.wp.Tape.zero() zeroes gradients passed via the grads parameter in wp.Tape.backward()
(GH-407).wp.array() not respecting the target dtype and shape when the given data is an another array with a CUDA interface
(GH-363).ImportError exception being thrown during interpreter shutdown on Windows when using the OpenGL renderer
(GH-412).AttributeError crash in the OpenGL renderer when moving the camera (GH-426).wp.sim.ModelBuilder default parameters (GH-429).wp.tile_extract() when the block dimension is smaller than the tile size.wp.autograd.jacobian() where in some cases gradients were not zeroed-out properly.wp.autograd.jacobian_plot().len() operator returning the total size of a matrix instead of its first dimension.wp.sim.SemiImplicitIntegrator and
wp.sim.FeatherstoneIntegrator (GH-349).up_axis, color in OpenGLRenderer (GH-448).Add PyTorch basics and custom operators notebooks to the notebooks directory.
notebooks directory.wp.launch_tiled() not returning a Launch object when passed record_cmd=True.wp.func when called from Python's runtime
(GH-386).wp.atomic_add(), wp.atomic_sub(),
wp.atomic_max(), or wp.atomic_min() as being written to (GH-378)..meta files into Warp kernel cache on Windows.Support for cooperative tile-based primitives using cuBLASDx and cuFFTDx, please see the tile documentation for details.
reversed() built-in for iterators (GH-311)..nvdb files with the save_to_nvdb method.wp.fem.Trimesh3D and wp.fem.Quadmesh3D geometry types for 3D surfaces with new example_distortion_energy example."add" option to wp.fem.integrate() for accumulating integration result to existing output."assembly" option to wp.fem.integrate() for selecting between more memory-efficient or more
computationally efficient integration algorithms.curl and div operators, respectively.wp.sim.ModelBuilder now includes methods to color particles for use with wp.sim.VBDIntegrator(),
users should call builder.color() before finalizing assets.wp.sim.Model.particle_radius
array (docs), replacing the previous
hard-coded value of 0.01 (GH-329).particle_radius parameter to wp.sim.ModelBuilder.add_cloth_mesh() and wp.sim.ModelBuilder.add_cloth_grid()
to set a uniform radius for the added particles.wp.array attributes (GH-364).notebooks directory.wp.Int, wp.Float, and wp.Scalar generic annotation types to the public API.wp.fem.cells(), wp.fem.to_inner_cell(), wp.fem.to_outer_cell() operators.wp.randn() samples a normal distribution of mean 0 and variance 1.wp.printf() built-in.place setting of paddle backend.wp.expect_neq() overloads missing for scalar types.wp.kernel or a wp.func object is annotated to return a None value..nvdb files.wp.printf() erroring out when no variadic arguments are passed (GH-333).Make the output of wp.print() in backward kernels consistent for all supported data types.
wp.print() in backward kernels consistent for all supported data types.1.3.0).dtype values.Texture Write node, used in the Mandelbrot Omniverse sample, sometimes erroring out in multi-GPU environments.Re-introduced the wp.rand*(), wp.sample*(), and wp.poisson() onto the Python scope to revert a breaking change.
iter_reverse() not working as expected for ranges with steps other than 1 (GH-311).wp.sparse.BsrMatrix object is reused for storing matrices of different shapes.wp.fem.utils.symmetric_eigenvalues_qr.ModelBuilder.add_builder(builder) to correctly update articulation_start and thereby articulation_count when builder contains more than one articulation.wp.rand*(), wp.sample*(), and wp.poisson() onto the Python scope to revert a breaking change.Support for a new wp.static(expr) function that allows arbitrary Python expressions to be evaluated at the time of function/kernel definition (docs).
wp.static(expr) function that allows arbitrary Python expressions to be evaluated at the time of
function/kernel definition (docs).warp.fem (docs).wp.kernel and wp.func objects from within closures.wp.func.jax_kernel() (GH-310).wp.mod() for vector types (GH-282).% to Python's runtime scalar and vector types.atomic_add, atomic_max, and atomic_min (GH-284).q.w).omni.warp extension.warp.sim.VBDIntegrator now supports body-particle collision.wp.sim.Model.edge_indices now includes boundary edges.wp.rand*(), wp.sample*(), and wp.poisson() from the Python scope.wp.Mesh.points is now a property instead of a raw data member, its reference can be changed after the mesh is initialized.if/else/elif statements with constant conditions are resolved at compile time with no branches being inserted in the generated code.warp.fem.wp.func erroring out when defining a Tuple as a return type hint (GH-302).+=, -=) adjoints to compute gradients correctly in the backwards passv[1] = x--num_tiles 1 in example_render_opengl.py (GH-306).FeatherstoneIntegrator when bodies and particles collide.FeatherstoneIntegrator where eval_rigid_jacobian could give incorrect results or reach an infinite
loop when the body and joint indices were not in the same order. Added Model.joint_ancestor to fix the indexing
from a joint to its parent joint in the articulation.add_edges() called from ModelBuilder.add_cloth_mesh() (GH-319).compute-sanitizer initcheck tool when using wp.Mesh.stream argument in array.__dlpack__().while statements.wp.Tape context.kDLBool instead of kDLUInt for DLPack interop of Booleans.Fix an aliasing issue with zero-copy array initialization from NumPy introduced in Warp 1.3.0.
wp.Volume.load_from_numpy() behavior when bg_value is a sequence of values.wp.svd3 with fp64 numbers (GH-281).wp.bvh_query_ray() where the direction instead of the reciprocal direction was used
(GH-288).wp.sim.collide.triangle_closest_point_barycentric() where the returned barycentric coordinates may be
incorrect when the closest point lies on an edge.np.int32.input_output_mask argument to autograd.jacobian and
autograd.jacobian_fd (GH-289).ModelBuilder.collapse_fixed_joints() to correctly update the body centers of mass and the
ModelBuilder.articulation_start array.wp.fem.ExplicitQuadrature (regression from 1.3.0).wp.bvh_query_aabb() returns parts that overlap the bounding volume.wp.synchronize() from PyTorch autograd function exampleTape.check_kernel_array_access() and Tape.reset_array_read_flags() are now private methods.Warp Core improvements
(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 a wp.ScopedTimer.wp.clear_kernel_cache() to the public API. This is equivalent to wp.build.clear_kernel_cache().wp.config variables.wp.matmul() CPU fallback to use dtype explicitly in np.matmul() callfrom __future__ import annotations (GH-256).wp.launch() directly via __cuda_array_interface__ and __array_interface__, up to 2.5x faster conversion from PyTorchreturn_ctype argument to wp.from_torch()wp.abs() and wp.sign() for vector typeswp.float16(1.23) * wp.float16(2.34))wp.copy(), wp.clone(), and array.assign() differentiability__new__() methods for all class __del__() methods to handle when a class instance is created but not instantiated before garbage collectionwp.quat_t suffix: wp.BVHQuery, wp.HashGridQuery, wp.MeshQueryAABB, wp.MeshQueryPoint, and wp.MeshQueryRaywp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)warp.autograd improvements:
warp.autograd module with utility functions gradcheck(), jacobian(), and jacobian_fd() for debugging kernel Jacobians (docs)wp.config.verify_autograd_array_access is true in-place operations on arrays on the Tape that could break gradient computation will be detected (docs)@wp.func_replay functions and native snippets would not trigger module recompilationwarp.sim improvements:
self.rigid_mesh_contact_max is zero (default behavior).mask argument to wp.sim.eval_fk() now accepts both integer and boolean arrays to mask articulations.ModelBuilder.joint_act in ModelBuilder.collapse_fixed_joints() (affected floating-base systems)ModelBuilder.plot_articulation() to visualize the articulation tree of a rigid-body mechanism__new__() method (missing instance return and *args parameter)upaxis variable in ModelBuilder and the rendering thereof in OpenGLRendererwarp.sparse improvements:
bsr_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)A += 0.5 * B, y = x @ Cwarp.fem new features and fixes:
wp.fem.lookup() operator now supports wp.fem.Tetmesh and wp.fem.Trimesh2D geometrieswp.fem.Subdomain), free-slip boundary conditionswp.fem.UniformField, wp.fem.ImplicitField and wp.fem.NonconformingFieldstreamlines, magnetostatics and nonconforming_contact examples, updated mixed_elasticity to use a nonlinear modelwp.fem.PicQuadrature w.r.t. positions and measuresFix accuracy of 3x3 SVD wp.svd3 with fp64 numbers (GH-281).
wp.svd3 with fp64 numbers (GH-281).wp.bvh_query_ray() where the direction instead of the reciprocal direction was used
(GH-288).wp.sim.collide.triangle_closest_point_barycentric() where the returned barycentric coordinates may be
incorrect when the closest point lies on an edge.np.int32.input_output_mask argument to autograd.jacobian and
autograd.jacobian_fd (GH-289).ModelBuilder.collapse_fixed_joints() to correctly update the body centers of mass and the
ModelBuilder.articulation_start array.wp.fem.ExplicitQuadrature (regression from 1.3.0).wp.bvh_query_aabb() returns parts that overlap the bounding volume.wp.synchronize() from PyTorch autograd function exampleTape.check_kernel_array_access() and Tape.reset_array_read_flags() are now private methods.(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 a wp.ScopedTimer.wp.clear_kernel_cache() to the public API. This is equivalent to wp.build.clear_kernel_cache().wp.config variables.wp.matmul() CPU fallback to use dtype explicitly in np.matmul() callfrom __future__ import annotations (GH-256).wp.launch() directly via __cuda_array_interface__ and __array_interface__, up to 2.5x faster conversion from PyTorchreturn_ctype argument to wp.from_torch()wp.abs() and wp.sign() for vector typeswp.float16(1.23) * wp.float16(2.34))wp.copy(), wp.clone(), and array.assign() differentiability__new__() methods for all class __del__() methods to handle when a class instance is created but not instantiated before garbage collectionwp.quat_t suffix: wp.BVHQuery, wp.HashGridQuery, wp.MeshQueryAABB, wp.MeshQueryPoint, and wp.MeshQueryRaywp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)Remove wp.synchronize() from PyTorch autograd function example
wp.synchronize() from PyTorch autograd function exampleTape.check_kernel_array_access() and Tape.reset_array_read_flags() are now private methods.Warp Core improvements
(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 a wp.ScopedTimer.wp.clear_kernel_cache() to the public API. This is equivalent to wp.build.clear_kernel_cache().wp.config variables.wp.matmul() CPU fallback to use dtype explicitly in np.matmul() callfrom __future__ import annotations (GH-256).wp.launch() directly via __cuda_array_interface__ and __array_interface__, up to 2.5x faster conversion from PyTorchreturn_ctype argument to wp.from_torch()wp.abs() and wp.sign() for vector typeswp.float16(1.23) * wp.float16(2.34))wp.copy(), wp.clone(), and array.assign() differentiability__new__() methods for all class __del__() methods to handle when a class instance is created but not instantiated before garbage collectionwp.quat_t suffix: wp.BVHQuery, wp.HashGridQuery, wp.MeshQueryAABB, wp.MeshQueryPoint, and wp.MeshQueryRaywp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)warp.autograd improvements:
warp.autograd module with utility functions gradcheck(), jacobian(), and jacobian_fd() for debugging kernel Jacobians (docs)wp.config.verify_autograd_array_access is true in-place operations on arrays on the Tape that could break gradient computation will be detected (docs)@wp.func_replay functions and native snippets would not trigger module recompilationwarp.sim improvements:
self.rigid_mesh_contact_max is zero (default behavior).mask argument to wp.sim.eval_fk() now accepts both integer and boolean arrays to mask articulations.ModelBuilder.joint_act in ModelBuilder.collapse_fixed_joints() (affected floating-base systems)ModelBuilder.plot_articulation() to visualize the articulation tree of a rigid-body mechanism__new__() method (missing instance return and *args parameter)upaxis variable in ModelBuilder and the rendering thereof in OpenGLRendererwarp.sparse improvements:
bsr_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)A += 0.5 * B, y = x @ Cwarp.fem new features and fixes:
wp.fem.lookup() operator now supports wp.fem.Tetmesh and wp.fem.Trimesh2D geometrieswp.fem.Subdomain), free-slip boundary conditionswp.fem.UniformField, wp.fem.ImplicitField and wp.fem.NonconformingFieldstreamlines, magnetostatics and nonconforming_contact examples, updated mixed_elasticity to use a nonlinear modelwp.fem.PicQuadrature w.r.t. positions and measuresUpdate to CUDA 12.x by default (requires NVIDIA driver 525 or newer), please see README.md for commands to install CUDA 11.x binaries for older driver
Warp Core improvements
(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 a wp.ScopedTimer.wp.clear_kernel_cache() to the public API. This is equivalent to wp.build.clear_kernel_cache().wp.config variables.wp.matmul() CPU fallback to use dtype explicitly in np.matmul() callfrom __future__ import annotations (GH-256).wp.launch() directly via __cuda_array_interface__ and __array_interface__, up to 2.5x faster conversion from PyTorchreturn_ctype argument to wp.from_torch()wp.abs() and wp.sign() for vector typeswp.float16(1.23) * wp.float16(2.34))wp.copy(), wp.clone(), and array.assign() differentiability__new__() methods for all class __del__() methods to handle when a class instance is created but not instantiated before garbage collectionwp.quat_t suffix: wp.BVHQuery, wp.HashGridQuery, wp.MeshQueryAABB, wp.MeshQueryPoint, and wp.MeshQueryRaywp.array(ptr=...) to allow initializing arrays from pointer addresses inside of kernels (GH-206)warp.autograd improvements:
warp.autograd module with utility functions gradcheck(), jacobian(), and jacobian_fd() for debugging kernel Jacobians (docs)wp.config.verify_autograd_array_access is true in-place operations on arrays on the Tape that could break gradient computation will be detected (docs)@wp.func_replay functions and native snippets would not trigger module recompilationwarp.sim improvements:
self.rigid_mesh_contact_max is zero (default behavior).mask argument to wp.sim.eval_fk() now accepts both integer and boolean arrays to mask articulations.ModelBuilder.joint_act in ModelBuilder.collapse_fixed_joints() (affected floating-base systems)ModelBuilder.plot_articulation() to visualize the articulation tree of a rigid-body mechanism__new__() method (missing instance return and *args parameter)upaxis variable in ModelBuilder and the rendering thereof in OpenGLRendererwarp.sparse improvements:
bsr_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)A += 0.5 * B, y = x @ Cwarp.fem new features and fixes:
wp.fem.lookup() operator now supports wp.fem.Tetmesh and wp.fem.Trimesh2D geometrieswp.fem.Subdomain), free-slip boundary conditionswp.fem.UniformField, wp.fem.ImplicitField and wp.fem.NonconformingFieldstreamlines, magnetostatics and nonconforming_contact examples, updated mixed_elasticity to use a nonlinear modelwp.fem.PicQuadrature w.r.t. positions and measuresFix Warp not being initialized when constructing arrays with wp.array()
wp.array()wp.is_mempool_access_supported() not resolving the provided device arguments to wp.context.Devicewp.NAN or wp.nan.wp.isnan(), wp.isinf(), and wp.isfinite() for scalars, vectors, matrices, etc.wp.constant() variables declared in a Warp program.wp.MarchingCubes on field dimensions and sizewp.Mesh BVH (GH-225)sm_75 (from sm_70), enabling Turing ISA featureswarp.Volume):
wp.volume_lookup_index(), wp.volume_sample_index() and generic wp.volume_sample()/wp.volume_lookup()/wp.volume_store() kernel-level functionswarp.fem can now work directly on NanoVDB grids using warp.fem.Nanogridwp.volume_sample_v() and wp.volume_store_*() adjointswp.volume_store() from overwriting grid background valueswarp.femwp.render.OpenGLRenderer via pyglet.options["headless"] = Truewp.render.RegisteredGLBuffer can fall back to CPU-bound copying if CUDA/OpenGL interop is not availablewp.sparse.bsr_mm() by ~5x on benchmark problemsjoint_act arrayswp.sim.FeatherstoneIntegrator()--msvc_path in build scriptswp.copy() params to record dest and src offset parameters on wp.Tape()wp.randn() to ensure return values are finitebool types in generic kernelswp.copy(), wp.clone(), and array.assign() differentiability__new__() methods for all class __del__() methods to
handle when a class instance is created but not instantiated before garbage collection.wp.quatFix Warp not being initialized when constructing arrays with wp.array()
wp.array()wp.is_mempool_access_supported() not resolving the provided device arguments to wp.context.Devicewp.NAN or wp.nan.wp.isnan(), wp.isinf(), and wp.isfinite() for scalars, vectors, matrices, etc.wp.constant() variables declared in a Warp program.wp.MarchingCubes on field dimensions and sizewp.Mesh BVH (GH-225)sm_75 (from sm_70), enabling Turing ISA featureswarp.Volume):
wp.volume_lookup_index(), wp.volume_sample_index() and generic wp.volume_sample()/wp.volume_lookup()/wp.volume_store() kernel-level functionswarp.fem can now work directly on NanoVDB grids using warp.fem.Nanogridwp.volume_sample_v() and wp.volume_store_*() adjointswp.volume_store() from overwriting grid background valueswarp.femwp.render.OpenGLRenderer via pyglet.options["headless"] = Truewp.render.RegisteredGLBuffer can fall back to CPU-bound copying if CUDA/OpenGL interop is not availablewp.sparse.bsr_mm() by ~5x on benchmark problemsjoint_act arrayswp.sim.FeatherstoneIntegrator()--msvc_path in build scriptswp.copy() params to record dest and src offset parameters on wp.Tape()wp.randn() to ensure return values are finitebool types in generic kernelsAdd a not-a-number floating-point constant that can be used as wp.NAN or wp.nan.
wp.NAN or wp.nan.wp.isnan(), wp.isinf(), and wp.isfinite() for scalars, vectors, matrices, etc.wp.constant() variables declared in a Warp program.wp.MarchingCubes on field dimensions and sizewp.Mesh BVH (GH-225)sm_75 (from sm_70), enabling Turing ISA featureswarp.Volume):
wp.volume_lookup_index(), wp.volume_sample_index() and generic wp.volume_sample()/wp.volume_lookup()/wp.volume_store() kernel-level functionswarp.fem can now work directly on NanoVDB grids using warp.fem.Nanogridwp.volume_sample_v() and wp.volume_store_*() adjointswp.volume_store() from overwriting grid background valueswarp.femwp.render.OpenGLRenderer via pyglet.options["headless"] = Truewp.render.RegisteredGLBuffer can fall back to CPU-bound copying if CUDA/OpenGL interop is not availablewp.sparse.bsr_mm() by ~5x on benchmark problemsjoint_act arrayswp.sim.FeatherstoneIntegrator()--msvc_path in build scriptswp.copy() params to record dest and src offset parameters on wp.Tape()wp.randn() to ensure return values are finitebool types in generic kernelswp.init() is no longer required to be called explicitly and will be performed on first call to the APIomni.warp.core's startup timeSupport returning a value from @wp.func_native CUDA functions using type hints
@wp.func_native CUDA functions using type hintswp.sim.FeatherstoneIntegratorwp.sim.collide()wp.ScopedTimer()wp.Tape.visualize()__cuda_array_interface__ attributestruct.to(device) to migrate struct arrayswp.sim.Modelwp.HashGrid.build()Ruff for formatting and lintingwp.launch()wp.optim.linear.cr()import warp by eliminating raising any exceptionsa[i][j] vs. a[i, j]support_level entry to the configuration file of the extensionsMake examples runnable from any location
README.md examples USD locationexample_graph_capture.py descriptionDocument Device total_memory and free_memory
total_memory and free_memorypython -m warp.examples.browse for browsing the examples folderexamples/optim/example_walker.py samplewp.synchronize_event() for blocking the host thread until a recorded event completesstdout captureAdd FeatherstoneIntegrator which provides more stable simulation of articulated rigid body dynamics in generalized coordinates (State.joint_q and Stat
FeatherstoneIntegrator which provides more stable simulation of articulated rigid body dynamics in generalized coordinates (State.joint_q and State.joint_qd)warp.sim.Control struct to store control inputs for simulations (optional, by default the Model control inputs are used as before); integrators now have a different simulation signature: integrator.simulate(model: Model, state_in: State, state_out: State, dt: float, control: Control)joint_act can now behave in 3 modes: with joint_axis_mode set to JOINT_MODE_FORCE it behaves as a force/torque, with JOINT_MODE_VELOCITY it behaves as a velocity target, and with JOINT_MODE_POSITION it behaves as a position target; joint_target has been removedModel.shape_materials.ka which controls the contact distance at which the adhesive force is appliedwp.ScopedCaptureenable_backward warning for callablesAdd examples assets to the wheel packages
wp.config.quiet = TrueDeprecation of owner argument - use deleter to transfer ownership
examples directory under warp/python -m warp.tests --helptorch.autograd.function example + docsexample_graph_captureverify_fp causing compiler errors and support CPU kernelsmatmul to be called in CUDA graph capturewp.launch to support tuple argstest_femassert_np_equal when NaN's and tolerance are involvedwarp.config.max_unroll, fix custom gradient unrolling@wp.func_native(snippet, replay_snippet=replay_snippet)CUDA_PATH environment variable or --cuda_path build option are not usedwp.ones() to efficiently create one-initialized arrayswp.config.graph_capture_module_load_default to wp.config.enable_graph_capture_module_load_by_defaultwp.config.enable_mempools_at_init to enable pooled allocators during Warp initialization (if supported)wp.is_mempool_supported() - check if a device supports pooled allocatorswp.is_mempool_enabled(), wp.set_mempool_enabled() - enable or disable pooled allocators per devicewp.set_mempool_release_threshold(), wp.get_mempool_release_threshold() - configure memory pool release thresholdwp.is_peer_access_supported() - check if the memory of a device can be accessed by a peer devicewp.is_peer_access_enabled(), wp.set_peer_access_enabled() - manage peer access for memory allocated using default CUDA allocatorswp.is_mempool_access_supported() - check if the memory pool of a device can be accessed by a peer devicewp.is_mempool_access_enabled(), wp.set_mempool_access_enabled() - manage access for memory allocated using pooled CUDA allocatorswp.ScopedStream can synchronize with the previous stream on entry and/or exit (only sync on entry by default)wp.copy(), wp.launch(), wp.capture_launch())deleter argument when constructing arrays
owner argument - use deleter to transfer ownershipwp.zeros(), wp.full(), and more)wp.matmul() to always use the correct CUDA contextNoise Deform are deterministic across different Kit sessionsUpdate the license to *NVIDIA Software License*, allowing commercial use (see LICENSE.md)
LICENSE.md)CONTRIBUTING.md guidelines (for NVIDIA employees)snippet and adj_snippet strings to fix cachingbuild_docs.py on Windows.py extension to warp/tests/walkthrough_debugwp.bool usage in vector and matrix typesenable_backward setting is set to False upon calling wp.Tape.backward()wp.from_torch()Noise Deform node for OmniGraph that deforms points using a perlin/curl noiseShow deprecation warnings only once
pip install warp-lang.PACKAGING.md and resembling that of Python itself:
release-0.11 branch.public branch, previously used to merge releases into and corresponding with the GitHub main branch, is retired.try/finallywp.optim.linear.cg, wp.optim.linear.bicgstab, wp.optim.linear.gmres, and wp.optim.linear.LinearOperatorfloat(wp.float32(1.23)) or ctypes.c_float(wp.float32(1.23))wp.infexamples/example_diffsim_mass_spring_cage.py-s, --suite option for only running tests belonging to the given suiteswp.render.OpenGLRendererwp.mesh_query_ray()Remove deprecated wp.ScopedCudaGuard(), please use wp.ScopedDevice() instead
warp.array.reshape() to handle -1 dimensionswp.sim.create_soft_body_contacts()warp.from_torch(), warp.to_torch() plus documentationwp.mesh_query_point_nosign() - closest point query with no sign determinationwp.mesh_query_point_sign_normal() - closest point query with sign from angle-weighted normalwp.mesh_query_point_sign_winding_number() - closest point query with fast winding number sign determinationwarp.sparse module:
wp.sparse.BsrMatrixwp.sparse.bsr_zeros(), wp.sparse.bsr_set_from_triplets() for constructionwp.sparse.bsr_mm(), wp.sparse_bsr_mv() for matrix-matrix and matrix-vector products respectivelywp.utils.array_scan() - prefix sum (inclusive or exlusive)wp.utils.array_sum() - sum across arraywp.utils.radix_sort_pairs() - in-place radix sort (key,value) pairs@wp.func functions from Python (outside of kernel scope)wp.Launch object that can be replayed with low overhead, use wp.launch(..., record_cmd=True) to generate a command objectwp.struct kernel arguments, up to 20x faster launches for kernels with large structs or number of paramsomni.warp.nodes module
omni.warp.nodes.mesh_create_bundle(), omni.warp.nodes.mesh_get_points(), etc.wp.array:
wp.empty() no longer zeroes-out memory and returns an uninitialized array, as intendedarray.zero_() and array.fill_() work with non-contiguous arraysarray.fill_() can now take lists or other sequences when filling arrays of vectors or matrices, e.g. arr.fill_([[1, 2], [3, 4]])array.fill_() now works with arrays of structs (pass a struct instance)wp.copy() gracefully handles copying between non-contiguous arrays on different deviceswp.full() and wp.full_like(), e.g., a = wp.full(shape, value)device argument to wp.empty_like(), wp.zeros_like(), wp.full_like(), and wp.clone()indexedarray methods .zero_(), .fill_(), and .assign()indexedarray methods .numpy() and .list()array.list() to work with arrays of any Warp data typearray.list() synchronization issue with CUDA arraysarray.numpy() called on an array of structs returns a structured NumPy array with named fieldsError: No module named 'omni.warp.core' when running some Kit configurations (e.g.: stubgen)wp.struct instance address being included in module content hashwp.BVH.refit() when executed on the CPUwp.struct constructorwp.float16 vectors and matrices in Pythonwp.float16 members in structswp.ScopedCudaGuard(), please use wp.ScopedDevice() insteadwp.simwp.array.reshape() to handle -1 dimensionswp.sim.create_soft_body_contacts()wp.from_torch(), wp.to_torch() plus documentationDeprecate wp.Model.soft_contact_distance which is now replaced by wp.Model.particle_radius
wp.verbose if using gradients)wp.breakpoint(), and test_debug.py@wp.func functionspass, continue, and break statements__sincos_stret symbol for macOSwp.Mesh.points, and other cases where arrays are passed to native functions@ operator as an alias for wp.matmul()ModelBuilder.add_particle has a new radius argument, Model.particle_radius is now a Warp arrayModel.particle_flags Warp array, introduce PARTICLE_FLAG_ACTIVE to define whether a particle is being simulated and participates in contact dynamics&, |, ~, <<, >>cpu devicesomni.warp into omni.warp.core for Omniverse applications that want to use the Warp Python module with minimal additional dependencies@wp.struct registration during hot reloadunot() operator so kernel writers can use if not array: syntaxwp.hash_grid_point_id() now returns -1 if the wp.HashGrid has not been reserved beforewp.Model.soft_contact_distance which is now replaced by wp.Model.particle_radiusAdd Texture Write node for updating dynamic RTX textures from Warp kernels / nodes
Texture Write node for updating dynamic RTX textures from Warp kernels / nodeswp.load_module() to pre-load specific modules (pass recursive=True to load recursively)wp.poisson() for sampling Poisson distributionswp.sim.parse_usd()--standalone build optionwp.ScopedTimer()matrix.get_row(), matrix.set_row()wp.struct types within kernelsslice = array[indices] will now generate a sparse slice of array datadef compute(param: Any):with wp.ScopedDevice("cuda") as device: syntax (same for wp.ScopedStream(), wp.Tape())wp.vector(), and wp.matrix()I = wp.identity(n=3, dtype=float)wp.pos())wp.constant variables to be used directly in Python without having to use .val memberwp.struct typeswp.struct from functions--quick build for faster local dev. iteration (uses a reduced set of SASS arches)requires_grad parameter to wp.from_torch() to override gradient allocationwp.Tape()wp.matmul() with tape backward passwp.matmul()wp.launch(), up to 3x faster launches in common caseswp.randf() conversion to float to reduce bias for uniform samplingwp.func and wp.constant types from inside Python closureswp.constant variables can now be treated as their true type, accessing the underlying value through constant.val is no longer supportedwp.sim.model.ground_plane is now a wp.array to support gradient, users should call builder.set_ground_plane() to create the groundwp.sim capsule, cones, and cylinders are now aligned with the default USD up-axisReduce test time for vec/math types
Updated wp.from_torch() to support more data types
wp.from_torch() to support more data typeswp.from_torch() to automatically determine the target Warp data type if not specifiedwp.from_torch() to support non-contiguous tensors with arbitrary strideswp.matmul() and wp.matmul_batched()mat33 types, see wp.qr3(), and wp.eig3()wp.config.quiet = True)@wp.kernel(enable_backward=False)imp package with importlibwp.quat_slerp())Fix strides computation in array_t constructor, fixes a bug with accessing mesh indices through mesh.indices[]
Add smoothed particle hydrodynamics (SPH) example, see example_sph.py
example_sph.pyarray.shape inside kernels, e.g.: width = arr.shape[0]wp.Bvh and bvh_query_ray(), bvh_query_aabb() functionsspatial_vector, spatial_matrix typeswp.lerp() and wp.smoothstep() builtinswp.optim module with implementation of the Adam optimizer for float and vector typeswp.length_sq(), wp.trace() for vector / matrix types respectivelywp.quat_rpy(), wp.determinant()wp.atomic_min(), wp.atomic_max() operatorswarp.sim.model.add_cloth_mesh()wp.Volume.allocate(), and wp.Volume.allocate_by_tiles()wp.volume_store_i(), wp.volume_store_f(), wp.volume_store_v()example_jacobian_ik.pywp.Tape.zero() support for wp.struct typeswp.Mesh object accept both 1d and 2d arrays of face vertex indicesimportlib.reload()wp.constants() not invalidating kernels.ptx versions are presentUpdate all samples to use GPU interop path by default
wp.sample_unit_cube(), wp.sample_unit_disk(), etcwp.lower_bound() for searching sorted arrayswp.set_module_options({"enable_backward": False}), True by default__file__ attributewp.func() definitionswp.config.mode == "debug", this enables bounds checking on CUDA kernels in debug moderuntime = None errors when reloading the Warp modulewp.config.ptx_target_arch__init__.py, defer them to individual functions that need themwp.mesh_query_point()wp.HashGrid memory leak when creating/destroying grids@wp.struct referenceswp.is_cuda_available() to check availabilitywp.render classFix for marching cubes reallocation after initialization
wp.closest_point_edge_edge() builtinwp.sim.ModelBuilder.add_cloth_mesh()wp.float16wp.array.strides@wp.struct decoratorwp.config.verify_fpwp.MarchingCubes classr = m33[i]wp.log2() and wp.log10() builtinswp.sim.ModelBuilder objects to improve env. creation performance for RLwp.Tape.zero()requires_grad=Truewp.mat22 constructor adjointwp.closest_point_edge_edge() builtinwp.sim.ModelBuilder.add_cloth_mesh()wp.set_device(), wp.get_device(), wp.ScopedDevice"cuda:0", "cuda:1"), see wp.get_cuda_devices(), wp.get_cuda_device_count(), wp.get_cuda_device()"cuda"wp.map_cuda_device(), wp.unmap_cuda_device()wp.device_from_torch(), wp.device_to_torch()wp.get_preferred_device())wp.ScopedCudaGuard is deprecated, use wp.ScopedDevice insteadwp.synchronize() now synchronizes all devices; for finer-grained control, use wp.synchronize_device()"cuda" now refers to the current CUDA context, rather than a specific device like "cuda:0" or "cuda:1"NumPy implementations of many builtin methods have been moved to warp.utils and will be deprecated
wp.constant changes not updating module hashwp.array.grad, users should take care to always call wp.Tape.zero() to clear gradients between different invocations of wp.Tape.backward()wp.array.fill_() to set all entries to a scalar value (4-byte values only currently)capture option has been removed, users can now capture tapes inside existing CUDA graphs (e.g.: inside Torch)requires_grad=True at creation timefrom import * inside Warp initializationwp.curlnoise()wp.from_torch() to correctly preserve shape@wp.func functions based on argument typewp.copy() for detailswp.mesh_get_point(), wp.mesh_get_index(), wp.mesh_eval_face_normal()wp.quat_identity() now call the Warp native implementation directly and will return a wp.quat object instead of NumPy arraywarp.utils and will be deprecated@wp.func functions should not be namespaced when called, e.g.: previously wp.myfunc() would work even if myfunc() was not a builtinwp.rpy2quat(), please use wp.quat_rpy() insteadFix for unrolling loops with negative bounds
hash_grid_build_device() not found when lib is compiled without CUDA supportwarp.dll not found on some Windows installations@wp.func functionsx = array[i,j,k] syntax to address a 3-dimensional arraylaunch(kernel, dim=(i,j,k), ... and i,j,k = wp.tid() to obtain thread indiceswp.config.mode = "debug" to enablewp.mlp()wp.volume_sample_i(), wp.volume_sample_f(), and wp.volume_sample_vec()wp.Volume.load_from_nvdb()wp.intersect_tri_tri()wp.ScopedTimer()wp.inverse(), wp.quat_to_matrix(), wp.quat_from_matrix()--fast-math) to kernel compilation by defaultwarp.torch import by default (if PyTorch is installed)wp.volume_sample_world() is now replaced by wp.volume_sample_f/i/vec() which operate in index (local) space. Users should use wp.volume_world_to_index() to transform points from world space to index space before sampling.wp.mlp() expects multi-dimensional arrays instead of one-dimensional arrays for inference, all other semantics remain the same as earlier versions of this API.wp.array.length member has been removed, please use wp.array.shape to access array dimensions, or use wp.array.size to get total element countdense_gemm(), dense_chol(), etc methods as experimental until we revisit themOne column per quarter.