warp-lang 1.6.1__py3-none-macosx_10_13_universal2.whl → 1.7.0__py3-none-macosx_10_13_universal2.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of warp-lang might be problematic. Click here for more details.
- warp/__init__.py +21 -7
- warp/autograd.py +14 -6
- warp/bin/libwarp-clang.dylib +0 -0
- warp/bin/libwarp.dylib +0 -0
- warp/build.py +424 -6
- warp/build_dll.py +20 -20
- warp/builtins.py +467 -368
- warp/codegen.py +193 -125
- warp/config.py +56 -12
- warp/constants.py +14 -6
- warp/context.py +524 -277
- warp/dlpack.py +22 -12
- warp/examples/__init__.py +14 -6
- warp/examples/assets/nonuniform.usd +0 -0
- warp/examples/assets/nvidia_logo.png +0 -0
- warp/examples/benchmarks/benchmark_api.py +14 -6
- warp/examples/benchmarks/benchmark_cloth.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_cupy.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_jax.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_numba.py +15 -0
- warp/examples/benchmarks/benchmark_cloth_numpy.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_paddle.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_pytorch.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_taichi.py +14 -6
- warp/examples/benchmarks/benchmark_cloth_warp.py +14 -6
- warp/examples/benchmarks/benchmark_gemm.py +82 -48
- warp/examples/benchmarks/benchmark_interop_paddle.py +14 -6
- warp/examples/benchmarks/benchmark_interop_torch.py +14 -6
- warp/examples/benchmarks/benchmark_launches.py +14 -6
- warp/examples/benchmarks/benchmark_tile_load_store.py +103 -0
- warp/examples/browse.py +14 -6
- warp/examples/core/example_cupy.py +14 -6
- warp/examples/core/example_dem.py +14 -6
- warp/examples/core/example_fluid.py +14 -6
- warp/examples/core/example_graph_capture.py +14 -6
- warp/examples/core/example_marching_cubes.py +14 -6
- warp/examples/core/example_mesh.py +14 -6
- warp/examples/core/example_mesh_intersect.py +14 -6
- warp/examples/core/example_nvdb.py +14 -6
- warp/examples/core/example_raycast.py +14 -6
- warp/examples/core/example_raymarch.py +14 -6
- warp/examples/core/example_render_opengl.py +14 -6
- warp/examples/core/example_sample_mesh.py +300 -0
- warp/examples/core/example_sph.py +14 -6
- warp/examples/core/example_torch.py +14 -6
- warp/examples/core/example_wave.py +14 -6
- warp/examples/fem/example_adaptive_grid.py +14 -6
- warp/examples/fem/example_apic_fluid.py +15 -7
- warp/examples/fem/example_burgers.py +16 -8
- warp/examples/fem/example_convection_diffusion.py +14 -6
- warp/examples/fem/example_convection_diffusion_dg.py +14 -6
- warp/examples/fem/example_deformed_geometry.py +15 -7
- warp/examples/fem/example_diffusion.py +14 -6
- warp/examples/fem/example_diffusion_3d.py +14 -6
- warp/examples/fem/example_diffusion_mgpu.py +14 -6
- warp/examples/fem/example_distortion_energy.py +15 -7
- warp/examples/fem/example_magnetostatics.py +20 -12
- warp/examples/fem/example_mixed_elasticity.py +14 -6
- warp/examples/fem/example_navier_stokes.py +14 -6
- warp/examples/fem/example_nonconforming_contact.py +14 -6
- warp/examples/fem/example_stokes.py +14 -6
- warp/examples/fem/example_stokes_transfer.py +14 -6
- warp/examples/fem/example_streamlines.py +14 -6
- warp/examples/fem/utils.py +24 -3
- warp/examples/interop/example_jax_callable.py +116 -0
- warp/examples/interop/example_jax_ffi_callback.py +132 -0
- warp/examples/interop/example_jax_kernel.py +205 -0
- warp/examples/optim/example_bounce.py +14 -6
- warp/examples/optim/example_cloth_throw.py +14 -6
- warp/examples/optim/example_diffray.py +14 -6
- warp/examples/optim/example_drone.py +14 -6
- warp/examples/optim/example_fluid_checkpoint.py +497 -0
- warp/examples/optim/example_inverse_kinematics.py +14 -6
- warp/examples/optim/example_inverse_kinematics_torch.py +14 -6
- warp/examples/optim/example_softbody_properties.py +14 -6
- warp/examples/optim/example_spring_cage.py +14 -6
- warp/examples/optim/example_trajectory.py +14 -6
- warp/examples/sim/example_cartpole.py +14 -6
- warp/examples/sim/example_cloth.py +14 -6
- warp/examples/sim/example_cloth_self_contact.py +14 -6
- warp/examples/sim/example_granular.py +14 -6
- warp/examples/sim/example_granular_collision_sdf.py +14 -6
- warp/examples/sim/example_jacobian_ik.py +14 -6
- warp/examples/sim/example_particle_chain.py +14 -6
- warp/examples/sim/example_quadruped.py +14 -6
- warp/examples/sim/example_rigid_chain.py +14 -6
- warp/examples/sim/example_rigid_contact.py +14 -6
- warp/examples/sim/example_rigid_force.py +14 -6
- warp/examples/sim/example_rigid_gyroscopic.py +14 -6
- warp/examples/sim/example_rigid_soft_contact.py +14 -6
- warp/examples/sim/example_soft_body.py +14 -6
- warp/examples/tile/example_tile_cholesky.py +14 -6
- warp/examples/tile/example_tile_convolution.py +14 -6
- warp/examples/tile/example_tile_fft.py +14 -6
- warp/examples/tile/example_tile_filtering.py +14 -6
- warp/examples/tile/example_tile_matmul.py +16 -10
- warp/examples/tile/example_tile_mlp.py +14 -6
- warp/examples/tile/example_tile_nbody.py +14 -6
- warp/examples/tile/example_tile_walker.py +14 -6
- warp/fabric.py +15 -0
- warp/fem/__init__.py +26 -1
- warp/fem/adaptivity.py +19 -4
- warp/fem/cache.py +15 -0
- warp/fem/dirichlet.py +15 -0
- warp/fem/domain.py +15 -0
- warp/fem/field/__init__.py +15 -0
- warp/fem/field/field.py +15 -0
- warp/fem/field/nodal_field.py +37 -68
- warp/fem/field/restriction.py +15 -0
- warp/fem/field/virtual.py +77 -23
- warp/fem/geometry/__init__.py +15 -0
- warp/fem/geometry/adaptive_nanogrid.py +24 -10
- warp/fem/geometry/closest_point.py +16 -1
- warp/fem/geometry/deformed_geometry.py +20 -2
- warp/fem/geometry/element.py +15 -0
- warp/fem/geometry/geometry.py +20 -0
- warp/fem/geometry/grid_2d.py +27 -12
- warp/fem/geometry/grid_3d.py +27 -15
- warp/fem/geometry/hexmesh.py +20 -7
- warp/fem/geometry/nanogrid.py +24 -11
- warp/fem/geometry/partition.py +15 -0
- warp/fem/geometry/quadmesh.py +28 -13
- warp/fem/geometry/tetmesh.py +18 -4
- warp/fem/geometry/trimesh.py +18 -8
- warp/fem/integrate.py +277 -93
- warp/fem/linalg.py +20 -5
- warp/fem/operator.py +15 -0
- warp/fem/polynomial.py +15 -0
- warp/fem/quadrature/__init__.py +15 -0
- warp/fem/quadrature/pic_quadrature.py +52 -22
- warp/fem/quadrature/quadrature.py +209 -25
- warp/fem/space/__init__.py +16 -1
- warp/fem/space/basis_function_space.py +19 -2
- warp/fem/space/basis_space.py +40 -18
- warp/fem/space/dof_mapper.py +15 -0
- warp/fem/space/function_space.py +15 -0
- warp/fem/space/grid_2d_function_space.py +15 -0
- warp/fem/space/grid_3d_function_space.py +15 -0
- warp/fem/space/hexmesh_function_space.py +17 -2
- warp/fem/space/nanogrid_function_space.py +15 -0
- warp/fem/space/partition.py +21 -2
- warp/fem/space/quadmesh_function_space.py +23 -8
- warp/fem/space/restriction.py +15 -0
- warp/fem/space/shape/__init__.py +15 -0
- warp/fem/space/shape/cube_shape_function.py +38 -23
- warp/fem/space/shape/shape_function.py +15 -0
- warp/fem/space/shape/square_shape_function.py +27 -12
- warp/fem/space/shape/tet_shape_function.py +15 -0
- warp/fem/space/shape/triangle_shape_function.py +16 -1
- warp/fem/space/tetmesh_function_space.py +18 -3
- warp/fem/space/topology.py +15 -0
- warp/fem/space/trimesh_function_space.py +17 -2
- warp/fem/types.py +15 -0
- warp/fem/utils.py +27 -6
- warp/jax.py +28 -7
- warp/jax_experimental/__init__.py +16 -0
- warp/{jax_experimental.py → jax_experimental/custom_call.py} +28 -33
- warp/jax_experimental/ffi.py +698 -0
- warp/jax_experimental/xla_ffi.py +602 -0
- warp/math.py +103 -6
- warp/native/array.h +28 -6
- warp/native/builtin.h +44 -9
- warp/native/bvh.cpp +18 -7
- warp/native/bvh.cu +57 -20
- warp/native/bvh.h +17 -7
- warp/native/clang/clang.cpp +45 -9
- warp/native/coloring.cpp +15 -6
- warp/native/crt.cpp +15 -6
- warp/native/crt.h +15 -6
- warp/native/cuda_crt.h +15 -6
- warp/native/cuda_util.cpp +29 -6
- warp/native/cuda_util.h +17 -6
- warp/native/error.cpp +15 -6
- warp/native/error.h +15 -6
- warp/native/exports.h +85 -63
- warp/native/fabric.h +15 -6
- warp/native/hashgrid.cpp +15 -6
- warp/native/hashgrid.cu +15 -6
- warp/native/hashgrid.h +15 -6
- warp/native/initializer_array.h +15 -6
- warp/native/intersect.h +41 -32
- warp/native/intersect_adj.h +48 -39
- warp/native/intersect_tri.h +17 -0
- warp/native/marching.cpp +16 -0
- warp/native/marching.cu +16 -7
- warp/native/marching.h +17 -0
- warp/native/mat.h +528 -15
- warp/native/mathdx.cpp +15 -6
- warp/native/matnn.h +15 -6
- warp/native/mesh.cpp +15 -6
- warp/native/mesh.cu +15 -6
- warp/native/mesh.h +25 -16
- warp/native/noise.h +15 -6
- warp/native/quat.h +114 -17
- warp/native/rand.h +21 -6
- warp/native/range.h +15 -6
- warp/native/reduce.cpp +15 -6
- warp/native/reduce.cu +15 -6
- warp/native/runlength_encode.cpp +15 -6
- warp/native/runlength_encode.cu +15 -6
- warp/native/scan.cpp +15 -6
- warp/native/scan.cu +15 -6
- warp/native/scan.h +15 -6
- warp/native/solid_angle.h +17 -0
- warp/native/sort.cpp +137 -65
- warp/native/sort.cu +167 -21
- warp/native/sort.h +23 -7
- warp/native/sparse.cpp +58 -28
- warp/native/sparse.cu +67 -23
- warp/native/spatial.h +15 -6
- warp/native/svd.h +131 -6
- warp/native/temp_buffer.h +15 -6
- warp/native/tile.h +316 -111
- warp/native/tile_reduce.h +61 -9
- warp/native/vec.h +83 -13
- warp/native/volume.cpp +100 -119
- warp/native/volume.cu +15 -6
- warp/native/volume.h +15 -6
- warp/native/volume_builder.cu +40 -16
- warp/native/volume_builder.h +21 -6
- warp/native/volume_impl.h +15 -6
- warp/native/warp.cpp +20 -12
- warp/native/warp.cu +114 -16
- warp/native/warp.h +34 -16
- warp/optim/__init__.py +14 -6
- warp/optim/adam.py +14 -6
- warp/optim/linear.py +25 -10
- warp/optim/sgd.py +14 -6
- warp/paddle.py +14 -6
- warp/render/__init__.py +14 -6
- warp/render/render_opengl.py +14 -6
- warp/render/render_usd.py +14 -6
- warp/render/utils.py +14 -6
- warp/sim/__init__.py +14 -7
- warp/sim/articulation.py +18 -10
- warp/sim/collide.py +35 -16
- warp/sim/graph_coloring.py +14 -6
- warp/sim/import_mjcf.py +463 -162
- warp/sim/import_snu.py +14 -7
- warp/sim/import_urdf.py +46 -18
- warp/sim/import_usd.py +14 -7
- warp/sim/inertia.py +14 -6
- warp/sim/integrator.py +14 -6
- warp/sim/integrator_euler.py +19 -11
- warp/sim/integrator_featherstone.py +17 -16
- warp/sim/integrator_vbd.py +222 -8
- warp/sim/integrator_xpbd.py +19 -11
- warp/sim/model.py +56 -19
- warp/sim/particles.py +14 -6
- warp/sim/render.py +14 -6
- warp/sim/utils.py +17 -2
- warp/sparse.py +657 -555
- warp/stubs.py +231 -19
- warp/tape.py +14 -6
- warp/tests/aux_test_class_kernel.py +14 -6
- warp/tests/aux_test_compile_consts_dummy.py +14 -6
- warp/tests/aux_test_conditional_unequal_types_kernels.py +14 -6
- warp/tests/aux_test_dependent.py +14 -6
- warp/tests/aux_test_grad_customs.py +14 -6
- warp/tests/aux_test_instancing_gc.py +14 -6
- warp/tests/aux_test_module_unload.py +14 -6
- warp/tests/aux_test_name_clash1.py +14 -6
- warp/tests/aux_test_name_clash2.py +14 -6
- warp/tests/aux_test_unresolved_func.py +14 -6
- warp/tests/aux_test_unresolved_symbol.py +14 -6
- warp/tests/cuda/__init__.py +0 -0
- warp/tests/{test_async.py → cuda/test_async.py} +14 -6
- warp/tests/{test_ipc.py → cuda/test_ipc.py} +14 -6
- warp/tests/{test_mempool.py → cuda/test_mempool.py} +53 -6
- warp/tests/{test_multigpu.py → cuda/test_multigpu.py} +14 -6
- warp/tests/{test_peer.py → cuda/test_peer.py} +14 -6
- warp/tests/{test_pinned.py → cuda/test_pinned.py} +14 -6
- warp/tests/{test_streams.py → cuda/test_streams.py} +85 -6
- warp/tests/geometry/__init__.py +0 -0
- warp/tests/{test_bvh.py → geometry/test_bvh.py} +14 -6
- warp/tests/{test_hash_grid.py → geometry/test_hash_grid.py} +14 -6
- warp/tests/{test_marching_cubes.py → geometry/test_marching_cubes.py} +14 -6
- warp/tests/{test_mesh.py → geometry/test_mesh.py} +14 -6
- warp/tests/{test_mesh_query_aabb.py → geometry/test_mesh_query_aabb.py} +14 -6
- warp/tests/{test_mesh_query_point.py → geometry/test_mesh_query_point.py} +80 -69
- warp/tests/{test_mesh_query_ray.py → geometry/test_mesh_query_ray.py} +15 -7
- warp/tests/{test_volume.py → geometry/test_volume.py} +55 -12
- warp/tests/{test_volume_write.py → geometry/test_volume_write.py} +14 -6
- warp/tests/interop/__init__.py +0 -0
- warp/tests/{test_dlpack.py → interop/test_dlpack.py} +42 -11
- warp/tests/{test_jax.py → interop/test_jax.py} +14 -6
- warp/tests/{test_paddle.py → interop/test_paddle.py} +14 -6
- warp/tests/{test_torch.py → interop/test_torch.py} +14 -6
- warp/tests/run_coverage_serial.py +14 -6
- warp/tests/sim/__init__.py +0 -0
- warp/tests/{disabled_kinematics.py → sim/disabled_kinematics.py} +23 -16
- warp/tests/{flaky_test_sim_grad.py → sim/flaky_test_sim_grad.py} +14 -6
- warp/tests/{test_collision.py → sim/test_collision.py} +16 -8
- warp/tests/{test_coloring.py → sim/test_coloring.py} +14 -7
- warp/tests/{test_model.py → sim/test_model.py} +55 -7
- warp/tests/{test_sim_grad_bounce_linear.py → sim/test_sim_grad_bounce_linear.py} +14 -6
- warp/tests/{test_sim_kinematics.py → sim/test_sim_kinematics.py} +16 -7
- warp/tests/sim/test_vbd.py +597 -0
- warp/tests/test_adam.py +14 -6
- warp/tests/test_arithmetic.py +14 -6
- warp/tests/test_array.py +14 -6
- warp/tests/test_array_reduce.py +14 -6
- warp/tests/test_assert.py +14 -6
- warp/tests/test_atomic.py +14 -6
- warp/tests/test_bool.py +15 -7
- warp/tests/test_builtins_resolution.py +14 -6
- warp/tests/test_closest_point_edge_edge.py +14 -6
- warp/tests/test_codegen.py +14 -6
- warp/tests/test_codegen_instancing.py +14 -6
- warp/tests/test_compile_consts.py +14 -6
- warp/tests/test_conditional.py +14 -6
- warp/tests/test_context.py +14 -6
- warp/tests/test_copy.py +14 -6
- warp/tests/test_ctypes.py +14 -6
- warp/tests/test_dense.py +14 -6
- warp/tests/test_devices.py +14 -6
- warp/tests/test_examples.py +42 -42
- warp/tests/test_fabricarray.py +14 -6
- warp/tests/test_fast_math.py +14 -6
- warp/tests/test_fem.py +37 -10
- warp/tests/test_fp16.py +14 -6
- warp/tests/test_func.py +14 -6
- warp/tests/test_future_annotations.py +14 -6
- warp/tests/test_generics.py +14 -6
- warp/tests/test_grad.py +14 -6
- warp/tests/test_grad_customs.py +14 -6
- warp/tests/test_grad_debug.py +14 -6
- warp/tests/test_implicit_init.py +14 -6
- warp/tests/test_import.py +14 -6
- warp/tests/test_indexedarray.py +14 -6
- warp/tests/test_intersect.py +14 -6
- warp/tests/test_iter.py +14 -6
- warp/tests/test_large.py +14 -6
- warp/tests/test_launch.py +14 -6
- warp/tests/test_lerp.py +14 -6
- warp/tests/test_linear_solvers.py +15 -11
- warp/tests/test_lvalue.py +14 -6
- warp/tests/test_mat.py +247 -85
- warp/tests/test_mat_lite.py +14 -6
- warp/tests/test_mat_scalar_ops.py +18 -10
- warp/tests/test_math.py +14 -6
- warp/tests/test_mlp.py +14 -6
- warp/tests/test_module_hashing.py +14 -6
- warp/tests/test_modules_lite.py +14 -6
- warp/tests/test_noise.py +14 -6
- warp/tests/test_operators.py +14 -6
- warp/tests/test_options.py +14 -6
- warp/tests/test_overwrite.py +15 -60
- warp/tests/test_print.py +14 -6
- warp/tests/test_quat.py +81 -52
- warp/tests/test_rand.py +58 -43
- warp/tests/test_reload.py +14 -6
- warp/tests/test_rounding.py +14 -6
- warp/tests/test_runlength_encode.py +14 -6
- warp/tests/test_scalar_ops.py +14 -6
- warp/tests/test_smoothstep.py +14 -6
- warp/tests/test_snippet.py +15 -0
- warp/tests/test_sparse.py +61 -12
- warp/tests/test_spatial.py +89 -6
- warp/tests/test_special_values.py +14 -6
- warp/tests/test_static.py +15 -7
- warp/tests/test_struct.py +14 -6
- warp/tests/test_tape.py +14 -6
- warp/tests/test_transient_module.py +14 -6
- warp/tests/test_triangle_closest_point.py +14 -6
- warp/tests/test_types.py +14 -6
- warp/tests/test_utils.py +98 -10
- warp/tests/test_vec.py +60 -40
- warp/tests/test_vec_lite.py +14 -6
- warp/tests/test_vec_scalar_ops.py +14 -6
- warp/tests/test_verify_fp.py +14 -6
- warp/tests/tile/__init__.py +0 -0
- warp/tests/{test_tile.py → tile/test_tile.py} +150 -57
- warp/tests/{test_tile_load.py → tile/test_tile_load.py} +15 -7
- warp/tests/{test_tile_mathdx.py → tile/test_tile_mathdx.py} +23 -12
- warp/tests/{test_tile_mlp.py → tile/test_tile_mlp.py} +39 -20
- warp/tests/{test_tile_reduce.py → tile/test_tile_reduce.py} +74 -7
- warp/tests/{test_tile_shared_memory.py → tile/test_tile_shared_memory.py} +14 -6
- warp/tests/{test_tile_view.py → tile/test_tile_view.py} +15 -7
- warp/tests/unittest_serial.py +15 -6
- warp/tests/unittest_suites.py +59 -65
- warp/tests/unittest_utils.py +16 -7
- warp/tests/walkthrough_debug.py +14 -6
- warp/thirdparty/unittest_parallel.py +15 -8
- warp/torch.py +14 -6
- warp/types.py +124 -664
- warp/utils.py +151 -78
- {warp_lang-1.6.1.dist-info → warp_lang-1.7.0.dist-info}/METADATA +39 -12
- warp_lang-1.7.0.dist-info/RECORD +429 -0
- {warp_lang-1.6.1.dist-info → warp_lang-1.7.0.dist-info}/WHEEL +1 -1
- warp_lang-1.7.0.dist-info/licenses/LICENSE.md +202 -0
- warp/examples/optim/example_walker.py +0 -309
- warp/native/cutlass_gemm.cpp +0 -34
- warp/native/cutlass_gemm.cu +0 -373
- warp/tests/test_matmul.py +0 -503
- warp/tests/test_matmul_lite.py +0 -403
- warp/tests/test_vbd.py +0 -378
- warp/tests/unused_test_misc.py +0 -69
- warp_lang-1.6.1.dist-info/LICENSE.md +0 -126
- warp_lang-1.6.1.dist-info/RECORD +0 -419
- {warp_lang-1.6.1.dist-info → warp_lang-1.7.0.dist-info}/top_level.txt +0 -0
|
@@ -0,0 +1,103 @@
|
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
15
|
+
|
|
16
|
+
import numpy as np
|
|
17
|
+
|
|
18
|
+
import warp as wp
|
|
19
|
+
|
|
20
|
+
BLOCK_DIM = 128
|
|
21
|
+
|
|
22
|
+
TILE = 32
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
def create_test_kernel(storage_type: str):
|
|
26
|
+
@wp.kernel
|
|
27
|
+
def load_store(a: wp.array2d(dtype=wp.float32), b: wp.array2d(dtype=wp.float32)):
|
|
28
|
+
i, j = wp.tid()
|
|
29
|
+
|
|
30
|
+
if wp.static(storage_type == "shared"):
|
|
31
|
+
a_tile = wp.tile_load(a, shape=(TILE, TILE), offset=(i * TILE, j * TILE), storage="shared")
|
|
32
|
+
else:
|
|
33
|
+
a_tile = wp.tile_load(a, shape=(TILE, TILE), offset=(i * TILE, j * TILE), storage="register")
|
|
34
|
+
|
|
35
|
+
wp.tile_store(b, a_tile, offset=(i * TILE, j * TILE))
|
|
36
|
+
|
|
37
|
+
return load_store
|
|
38
|
+
|
|
39
|
+
|
|
40
|
+
if __name__ == "__main__":
|
|
41
|
+
wp.config.quiet = True
|
|
42
|
+
wp.init()
|
|
43
|
+
wp.clear_kernel_cache()
|
|
44
|
+
wp.set_module_options({"fast_math": True, "enable_backward": False})
|
|
45
|
+
|
|
46
|
+
iterations = 100
|
|
47
|
+
rng = np.random.default_rng(42)
|
|
48
|
+
|
|
49
|
+
shared_benchmark_data = {}
|
|
50
|
+
register_benchmark_data = {}
|
|
51
|
+
memcpy_benchmark_data = {}
|
|
52
|
+
|
|
53
|
+
sizes = list(range(128, 4097, 128))
|
|
54
|
+
|
|
55
|
+
print(f"{'Transfer Size (Bytes)':<23s} {'Shared (GiB/s)':<16s} {'Register (GiB/s)':<18s} {'memcpy (GiB/s)':<16s}")
|
|
56
|
+
print("-" * 79)
|
|
57
|
+
|
|
58
|
+
for size in sizes:
|
|
59
|
+
a = wp.array(rng.random((size, size), dtype=np.float32), dtype=wp.float32)
|
|
60
|
+
b = wp.empty_like(a)
|
|
61
|
+
|
|
62
|
+
for storage_type in ("shared", "register"):
|
|
63
|
+
load_store = create_test_kernel(storage_type)
|
|
64
|
+
|
|
65
|
+
cmd = wp.launch_tiled(
|
|
66
|
+
load_store,
|
|
67
|
+
dim=(a.shape[0] // TILE, a.shape[1] // TILE),
|
|
68
|
+
inputs=[a],
|
|
69
|
+
outputs=[b],
|
|
70
|
+
block_dim=BLOCK_DIM,
|
|
71
|
+
record_cmd=True,
|
|
72
|
+
)
|
|
73
|
+
# Warmup
|
|
74
|
+
for _ in range(5):
|
|
75
|
+
cmd.launch()
|
|
76
|
+
|
|
77
|
+
with wp.ScopedTimer("benchmark", cuda_filter=wp.TIMING_KERNEL, print=False, synchronize=True) as timer:
|
|
78
|
+
for _ in range(iterations):
|
|
79
|
+
cmd.launch()
|
|
80
|
+
|
|
81
|
+
np.testing.assert_equal(a.numpy(), b.numpy())
|
|
82
|
+
|
|
83
|
+
timing_results = [result.elapsed for result in timer.timing_results]
|
|
84
|
+
avg_bw = 2.0 * (a.capacity / (1024 * 1024 * 1024)) / (1e-3 * np.mean(timing_results))
|
|
85
|
+
|
|
86
|
+
if storage_type == "shared":
|
|
87
|
+
shared_benchmark_data[a.capacity] = avg_bw
|
|
88
|
+
else:
|
|
89
|
+
register_benchmark_data[a.capacity] = avg_bw
|
|
90
|
+
|
|
91
|
+
# Compare with memcpy
|
|
92
|
+
with wp.ScopedTimer("benchmark", cuda_filter=wp.TIMING_MEMCPY, print=False, synchronize=True) as timer:
|
|
93
|
+
for _ in range(iterations):
|
|
94
|
+
wp.copy(b, a)
|
|
95
|
+
|
|
96
|
+
timing_results = [result.elapsed for result in timer.timing_results]
|
|
97
|
+
avg_bw = 2.0 * (a.capacity / (1024 * 1024 * 1024)) / (1e-3 * np.mean(timing_results))
|
|
98
|
+
memcpy_benchmark_data[a.capacity] = avg_bw
|
|
99
|
+
|
|
100
|
+
# Print results
|
|
101
|
+
print(
|
|
102
|
+
f"{a.capacity:<23d} {shared_benchmark_data[a.capacity]:<#16.4g} {register_benchmark_data[a.capacity]:<#18.4g} {memcpy_benchmark_data[a.capacity]:<#16.4g}"
|
|
103
|
+
)
|
warp/examples/browse.py
CHANGED
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2024 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
import os
|
|
9
17
|
import subprocess
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2024 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example CuPy
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example DEM
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example Fluid
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example Graph Capture
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example Marching Cubes
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example Mesh
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
#############################################################################
|
|
9
17
|
# Example Mesh Intersection
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example NanoVDB
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
#############################################################################
|
|
9
17
|
# Example Ray Cast
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# Example Ray March
|
|
@@ -1,9 +1,17 @@
|
|
|
1
|
-
# Copyright (c) 2022 NVIDIA CORPORATION.
|
|
2
|
-
#
|
|
3
|
-
#
|
|
4
|
-
#
|
|
5
|
-
#
|
|
6
|
-
#
|
|
1
|
+
# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
2
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
3
|
+
#
|
|
4
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
5
|
+
# you may not use this file except in compliance with the License.
|
|
6
|
+
# You may obtain a copy of the License at
|
|
7
|
+
#
|
|
8
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
|
9
|
+
#
|
|
10
|
+
# Unless required by applicable law or agreed to in writing, software
|
|
11
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
12
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
13
|
+
# See the License for the specific language governing permissions and
|
|
14
|
+
# limitations under the License.
|
|
7
15
|
|
|
8
16
|
###########################################################################
|
|
9
17
|
# OpenGL renderer example
|