warp-lang 1.6.0__py3-none-manylinux2014_aarch64.whl → 1.6.2__py3-none-manylinux2014_aarch64.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 +14 -6
- warp/autograd.py +14 -6
- warp/bin/warp-clang.so +0 -0
- warp/bin/warp.so +0 -0
- warp/build.py +14 -6
- warp/build_dll.py +14 -6
- warp/builtins.py +16 -7
- warp/codegen.py +24 -9
- warp/config.py +79 -27
- warp/constants.py +14 -6
- warp/context.py +236 -71
- warp/dlpack.py +14 -6
- warp/examples/__init__.py +14 -6
- 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/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 +15 -7
- warp/examples/core/example_mesh.py +15 -7
- 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_sph.py +14 -6
- warp/examples/core/example_torch.py +14 -6
- warp/examples/core/example_wave.py +15 -7
- warp/examples/fem/example_adaptive_grid.py +14 -6
- warp/examples/fem/example_apic_fluid.py +14 -6
- warp/examples/fem/example_burgers.py +14 -6
- 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 +14 -6
- 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 +14 -6
- warp/examples/fem/example_magnetostatics.py +14 -6
- 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 +15 -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_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/optim/example_walker.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 +95 -33
- 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 +14 -6
- warp/examples/tile/example_tile_mlp.py +14 -6
- warp/examples/tile/example_tile_nbody.py +40 -21
- warp/examples/tile/example_tile_walker.py +14 -6
- warp/fabric.py +15 -0
- warp/fem/__init__.py +15 -0
- warp/fem/adaptivity.py +15 -0
- 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 +15 -0
- warp/fem/field/restriction.py +15 -0
- warp/fem/field/virtual.py +15 -0
- warp/fem/geometry/__init__.py +15 -0
- warp/fem/geometry/adaptive_nanogrid.py +15 -0
- warp/fem/geometry/closest_point.py +15 -0
- warp/fem/geometry/deformed_geometry.py +15 -0
- warp/fem/geometry/element.py +15 -0
- warp/fem/geometry/geometry.py +15 -0
- warp/fem/geometry/grid_2d.py +15 -0
- warp/fem/geometry/grid_3d.py +15 -0
- warp/fem/geometry/hexmesh.py +15 -0
- warp/fem/geometry/nanogrid.py +15 -0
- warp/fem/geometry/partition.py +15 -0
- warp/fem/geometry/quadmesh.py +15 -0
- warp/fem/geometry/tetmesh.py +15 -0
- warp/fem/geometry/trimesh.py +15 -0
- warp/fem/integrate.py +15 -0
- warp/fem/linalg.py +15 -0
- 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 +15 -0
- warp/fem/quadrature/quadrature.py +15 -0
- warp/fem/space/__init__.py +15 -0
- warp/fem/space/basis_function_space.py +15 -0
- warp/fem/space/basis_space.py +15 -0
- 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 +15 -0
- warp/fem/space/nanogrid_function_space.py +15 -0
- warp/fem/space/partition.py +15 -0
- warp/fem/space/quadmesh_function_space.py +15 -0
- warp/fem/space/restriction.py +15 -0
- warp/fem/space/shape/__init__.py +15 -0
- warp/fem/space/shape/cube_shape_function.py +15 -0
- warp/fem/space/shape/shape_function.py +15 -0
- warp/fem/space/shape/square_shape_function.py +15 -0
- warp/fem/space/shape/tet_shape_function.py +15 -0
- warp/fem/space/shape/triangle_shape_function.py +15 -0
- warp/fem/space/tetmesh_function_space.py +15 -0
- warp/fem/space/topology.py +15 -0
- warp/fem/space/trimesh_function_space.py +15 -0
- warp/fem/types.py +15 -0
- warp/fem/utils.py +15 -0
- warp/jax.py +14 -6
- warp/jax_experimental.py +14 -6
- warp/math.py +14 -6
- warp/native/array.h +15 -6
- warp/native/builtin.h +15 -6
- warp/native/bvh.cpp +15 -6
- warp/native/bvh.cu +15 -6
- warp/native/bvh.h +15 -6
- warp/native/clang/clang.cpp +16 -7
- warp/native/coloring.cpp +15 -6
- warp/native/crt.cpp +15 -6
- warp/native/crt.h +16 -6
- warp/native/cuda_crt.h +15 -6
- warp/native/cuda_util.cpp +15 -6
- warp/native/cuda_util.h +15 -6
- warp/native/cutlass_gemm.cpp +15 -6
- warp/native/cutlass_gemm.cu +16 -7
- warp/native/error.cpp +15 -6
- warp/native/error.h +15 -6
- warp/native/exports.h +17 -0
- 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 +15 -6
- warp/native/intersect_adj.h +15 -6
- warp/native/intersect_tri.h +17 -0
- warp/native/marching.cpp +16 -0
- warp/native/marching.cu +15 -6
- warp/native/marching.h +17 -0
- warp/native/mat.h +31 -9
- 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 +15 -6
- warp/native/noise.h +15 -6
- warp/native/quat.h +15 -6
- warp/native/rand.h +15 -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 +15 -6
- warp/native/sort.cu +15 -6
- warp/native/sort.h +15 -6
- warp/native/sparse.cpp +15 -6
- warp/native/sparse.cu +15 -6
- warp/native/spatial.h +15 -6
- warp/native/svd.h +15 -6
- warp/native/temp_buffer.h +15 -6
- warp/native/tile.h +27 -14
- warp/native/tile_reduce.h +15 -6
- warp/native/vec.h +15 -6
- warp/native/volume.cpp +15 -6
- warp/native/volume.cu +15 -6
- warp/native/volume.h +15 -6
- warp/native/volume_builder.cu +15 -6
- warp/native/volume_builder.h +15 -6
- warp/native/volume_impl.h +15 -6
- warp/native/warp.cpp +15 -6
- warp/native/warp.cu +15 -6
- warp/native/warp.h +15 -6
- warp/optim/__init__.py +14 -6
- warp/optim/adam.py +14 -6
- warp/optim/linear.py +15 -0
- warp/optim/sgd.py +14 -6
- warp/paddle.py +14 -6
- warp/render/__init__.py +14 -6
- warp/render/render_opengl.py +37 -21
- warp/render/render_usd.py +24 -8
- warp/render/utils.py +14 -6
- warp/sim/__init__.py +14 -7
- warp/sim/articulation.py +14 -6
- warp/sim/collide.py +43 -22
- warp/sim/graph_coloring.py +14 -6
- warp/sim/import_mjcf.py +14 -7
- warp/sim/import_snu.py +14 -7
- warp/sim/import_urdf.py +34 -11
- warp/sim/import_usd.py +14 -7
- warp/sim/inertia.py +14 -6
- warp/sim/integrator.py +14 -6
- warp/sim/integrator_euler.py +14 -6
- warp/sim/integrator_featherstone.py +18 -17
- warp/sim/integrator_vbd.py +15 -6
- warp/sim/integrator_xpbd.py +14 -6
- warp/sim/model.py +76 -65
- warp/sim/particles.py +14 -6
- warp/sim/render.py +16 -8
- warp/sim/utils.py +15 -0
- warp/sparse.py +15 -0
- warp/stubs.py +16 -1
- warp/tape.py +14 -6
- warp/tests/__main__.py +15 -0
- 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/disabled_kinematics.py +14 -6
- warp/tests/flaky_test_sim_grad.py +14 -6
- warp/tests/run_coverage_serial.py +14 -6
- warp/tests/test_adam.py +14 -6
- warp/tests/test_arithmetic.py +14 -6
- warp/tests/test_array.py +40 -6
- warp/tests/test_array_reduce.py +14 -6
- warp/tests/test_assert.py +14 -6
- warp/tests/test_async.py +14 -6
- warp/tests/test_atomic.py +14 -6
- warp/tests/test_bool.py +14 -6
- warp/tests/test_builtins_resolution.py +14 -6
- warp/tests/test_bvh.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_collision.py +20 -12
- warp/tests/test_coloring.py +14 -7
- 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_dlpack.py +14 -6
- warp/tests/test_examples.py +21 -7
- warp/tests/test_fabricarray.py +14 -6
- warp/tests/test_fast_math.py +14 -6
- warp/tests/test_fem.py +14 -6
- 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_hash_grid.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_ipc.py +14 -6
- warp/tests/test_iter.py +14 -6
- warp/tests/test_jax.py +14 -6
- warp/tests/test_large.py +14 -6
- warp/tests/test_launch.py +91 -32
- warp/tests/test_lerp.py +14 -6
- warp/tests/test_linear_solvers.py +15 -0
- warp/tests/test_lvalue.py +14 -6
- warp/tests/test_marching_cubes.py +14 -6
- warp/tests/test_mat.py +89 -7
- warp/tests/test_mat_lite.py +14 -6
- warp/tests/test_mat_scalar_ops.py +14 -6
- warp/tests/test_math.py +14 -6
- warp/tests/test_matmul.py +14 -6
- warp/tests/test_matmul_lite.py +14 -6
- warp/tests/test_mempool.py +14 -6
- warp/tests/test_mesh.py +14 -6
- warp/tests/test_mesh_query_aabb.py +14 -6
- warp/tests/test_mesh_query_point.py +14 -6
- warp/tests/test_mesh_query_ray.py +14 -6
- warp/tests/test_mlp.py +14 -6
- warp/tests/test_model.py +14 -6
- warp/tests/test_module_hashing.py +14 -6
- warp/tests/test_modules_lite.py +14 -6
- warp/tests/test_multigpu.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 +19 -3
- warp/tests/test_paddle.py +14 -6
- warp/tests/test_peer.py +14 -6
- warp/tests/test_pinned.py +14 -6
- warp/tests/test_print.py +14 -6
- warp/tests/test_quat.py +14 -6
- warp/tests/test_rand.py +14 -6
- 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_sim_grad_bounce_linear.py +14 -6
- warp/tests/test_sim_kinematics.py +14 -6
- warp/tests/test_smoothstep.py +14 -6
- warp/tests/test_snippet.py +15 -0
- warp/tests/test_sparse.py +14 -6
- warp/tests/test_spatial.py +14 -6
- warp/tests/test_special_values.py +14 -6
- warp/tests/test_static.py +14 -6
- warp/tests/test_streams.py +14 -6
- warp/tests/test_struct.py +14 -6
- warp/tests/test_tape.py +14 -6
- warp/tests/test_tile.py +14 -6
- warp/tests/test_tile_load.py +58 -7
- warp/tests/test_tile_mathdx.py +14 -6
- warp/tests/test_tile_mlp.py +14 -6
- warp/tests/test_tile_reduce.py +14 -6
- warp/tests/test_tile_shared_memory.py +14 -6
- warp/tests/test_tile_view.py +14 -6
- warp/tests/test_torch.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 +14 -6
- warp/tests/test_vbd.py +14 -6
- warp/tests/test_vec.py +14 -6
- 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/test_volume.py +14 -6
- warp/tests/test_volume_write.py +14 -6
- warp/tests/unittest_serial.py +14 -6
- warp/tests/unittest_suites.py +14 -6
- warp/tests/unittest_utils.py +14 -6
- warp/tests/unused_test_misc.py +14 -6
- warp/tests/walkthrough_debug.py +14 -6
- warp/thirdparty/unittest_parallel.py +15 -7
- warp/torch.py +14 -6
- warp/types.py +80 -74
- warp/utils.py +14 -6
- warp_lang-1.6.2.dist-info/LICENSE.md +202 -0
- {warp_lang-1.6.0.dist-info → warp_lang-1.6.2.dist-info}/METADATA +44 -22
- warp_lang-1.6.2.dist-info/RECORD +419 -0
- {warp_lang-1.6.0.dist-info → warp_lang-1.6.2.dist-info}/WHEEL +1 -1
- warp_lang-1.6.0.dist-info/LICENSE.md +0 -126
- warp_lang-1.6.0.dist-info/RECORD +0 -419
- {warp_lang-1.6.0.dist-info → warp_lang-1.6.2.dist-info}/top_level.txt +0 -0
warp/native/mesh.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/noise.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/quat.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/rand.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
# pragma once
|
warp/native/range.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/reduce.cpp
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/reduce.cu
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "cuda_util.h"
|
warp/native/runlength_encode.cpp
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/runlength_encode.cu
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/scan.cpp
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "scan.h"
|
warp/native/scan.cu
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/scan.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/solid_angle.h
CHANGED
|
@@ -1,3 +1,20 @@
|
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
16
|
+
*/
|
|
17
|
+
|
|
1
18
|
// This code is adapted from https://github.com/alecjacobson/WindingNumber/tree/1e6081e52905575d8e98fb8b7c0921274a18752f
|
|
2
19
|
// The original license is below:
|
|
3
20
|
/*
|
warp/native/sort.cpp
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/sort.cu
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/sort.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/sparse.cpp
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "warp.h"
|
warp/native/sparse.cu
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#include "cuda_util.h"
|
warp/native/spatial.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/svd.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
// The MIT License (MIT)
|
warp/native/temp_buffer.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
warp/native/tile.h
CHANGED
|
@@ -1,9 +1,18 @@
|
|
|
1
|
-
|
|
2
|
-
* NVIDIA CORPORATION
|
|
3
|
-
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
6
|
-
*
|
|
1
|
+
/*
|
|
2
|
+
* SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
3
|
+
* SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
*
|
|
5
|
+
* Licensed under the Apache License, Version 2.0 (the "License");
|
|
6
|
+
* you may not use this file except in compliance with the License.
|
|
7
|
+
* You may obtain a copy of the License at
|
|
8
|
+
*
|
|
9
|
+
* http://www.apache.org/licenses/LICENSE-2.0
|
|
10
|
+
*
|
|
11
|
+
* Unless required by applicable law or agreed to in writing, software
|
|
12
|
+
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
13
|
+
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
14
|
+
* See the License for the specific language governing permissions and
|
|
15
|
+
* limitations under the License.
|
|
7
16
|
*/
|
|
8
17
|
|
|
9
18
|
#pragma once
|
|
@@ -1079,9 +1088,12 @@ struct tile_shared_t
|
|
|
1079
1088
|
constexpr bool contiguous_src = Layout::Stride::dim(lastdim) == 1;
|
|
1080
1089
|
const bool contiguous_dest = dest.data.strides[lastdim] == sizeof(T);
|
|
1081
1090
|
const int elements = (dest.data.shape[lastdim] - dest.offset[lastdim]);
|
|
1082
|
-
const bool
|
|
1091
|
+
const bool aligned_size = (elements*sizeof(T))%sizeof(float4) == 0;
|
|
1083
1092
|
|
|
1084
|
-
|
|
1093
|
+
float4* dest128 = (float4*)&dest.data.data[dest.index_from_coord(tile_coord(0,0))];
|
|
1094
|
+
const bool aligned_dst = (uint64_t)(dest128)%sizeof(float4) == 0;
|
|
1095
|
+
|
|
1096
|
+
if (contiguous_dest && contiguous_src && aligned_size && aligned_dst)
|
|
1085
1097
|
{
|
|
1086
1098
|
constexpr int M = Layout::Shape::dim(0);
|
|
1087
1099
|
constexpr int N = (Layout::Shape::dim(1)*sizeof(T))/sizeof(float4);
|
|
@@ -1089,10 +1101,9 @@ struct tile_shared_t
|
|
|
1089
1101
|
// alias of shared tile with 128bit type
|
|
1090
1102
|
using SrcLayout = tile_layout_strided_t<tile_shape_t<M, N>>;
|
|
1091
1103
|
tile_shared_t<float4, SrcLayout> src128((float4*)data.ptr);
|
|
1092
|
-
float4* dest128 = (float4*)&dest.data.data[dest.index_from_coord(tile_coord(0,0))];
|
|
1093
1104
|
|
|
1094
1105
|
assert(((uint64_t)(data.ptr))%sizeof(float4) == 0);
|
|
1095
|
-
assert(((uint64_t)(
|
|
1106
|
+
assert(((uint64_t)(dest128))%sizeof(float4) == 0);
|
|
1096
1107
|
|
|
1097
1108
|
const int stride_i = dest.data.strides[0]/sizeof(float4);
|
|
1098
1109
|
const int stride_j = 1;
|
|
@@ -1164,17 +1175,19 @@ struct tile_shared_t
|
|
|
1164
1175
|
constexpr bool contiguous_dest = Layout::Stride::dim(lastdim) == 1;
|
|
1165
1176
|
const bool contiguous_src = src.data.strides[lastdim] == sizeof(T);
|
|
1166
1177
|
const int elements = (src.data.shape[lastdim] - src.offset[lastdim]);
|
|
1167
|
-
const bool
|
|
1178
|
+
const bool aligned_size = (elements*sizeof(T))%sizeof(float4) == 0;
|
|
1179
|
+
|
|
1180
|
+
float4* src128 = (float4*)&src.data.data[src.index_from_coord(tile_coord(0,0))];
|
|
1181
|
+
const bool aligned_src = (uint64_t)(src128)%sizeof(float4) == 0;
|
|
1168
1182
|
|
|
1169
|
-
if (contiguous_dest && contiguous_src &&
|
|
1183
|
+
if (contiguous_dest && contiguous_src && aligned_size && aligned_src)
|
|
1170
1184
|
{
|
|
1171
1185
|
constexpr int M = Layout::Shape::dim(0);
|
|
1172
1186
|
constexpr int N = (Layout::Shape::dim(1)*sizeof(T))/sizeof(float4);
|
|
1173
1187
|
|
|
1174
1188
|
// alias of shared tile with 128bit type
|
|
1175
1189
|
using DestLayout = tile_layout_strided_t<tile_shape_t<M, N>>;
|
|
1176
|
-
tile_shared_t<float4, DestLayout> dest128((float4*)data.ptr);
|
|
1177
|
-
float4* src128 = (float4*)&src.data.data[src.index_from_coord(tile_coord(0,0))];
|
|
1190
|
+
tile_shared_t<float4, DestLayout> dest128((float4*)data.ptr);
|
|
1178
1191
|
|
|
1179
1192
|
assert(((uint64_t)(dest128.data.ptr))%sizeof(float4) == 0);
|
|
1180
1193
|
assert(((uint64_t)(src128))%sizeof(float4) == 0);
|