auvux-dsp 0.1.0.dev0__tar.gz

auvux_dsp-0.1.0.dev0/.clang-format

@@ -0,0 +1,6 @@
+BasedOnStyle: Google
+IndentWidth: 4
+ColumnLimit: 100
+AccessModifierOffset: -4
+DerivePointerAlignment: false
+PointerAlignment: Left

auvux_dsp-0.1.0.dev0/.clangd

@@ -0,0 +1,9 @@
+CompileFlags:
+  Add:
+    - -std=c++17
+    - -I/Users/pkiers/develop/auvux/auvux-dsp/src
+    - -I/Users/pkiers/develop/auvux/auvux-dsp/src/third_party
+    - -DAUVUX_HAVE_PFFFT=1
+    - -DAUVUX_HAVE_VDSP=1
+    - -I/Users/pkiers/develop/auvux/auvux-dsp/.venv/lib/python3.14/site-packages/pybind11/include
+    - -I/opt/homebrew/opt/python@3.14/Frameworks/Python.framework/Versions/3.14/include/python3.14

auvux_dsp-0.1.0.dev0/.env

@@ -0,0 +1 @@
+UV_PUBLISH_TOKEN=pypi-AgEIcHlwaS5vcmcCJDlhY2I4NWJjLWRmNzAtNGM3Ny04ZTdkLWRjYTQ0NzJmZTAzOAACKlszLCJiZjAzM2E3ZC1kYjE0LTRjNDAtYTk0Yy0zODM2YjlmY2I3MzciXQAABiDdf8s9ffSxqSmVcO6vg14IVDYXvlzWlXvv2rSgTTFgqg

auvux_dsp-0.1.0.dev0/.github/workflows/ci.yml

@@ -0,0 +1,71 @@
+name: ci
+
+# Manual-only until release readiness; restore the push/pull_request
+# triggers below to re-enable.
+#   push:
+#     branches: [main]
+#   pull_request:
+on:
+  workflow_dispatch:
+
+jobs:
+  test:
+    name: test (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install scikit-build-core pybind11 numpy pytest
+      - run: pip install --no-build-isolation -v .
+      - run: pytest tests -q
+
+  # Editable installs are the least-traveled path for namespace packages;
+  # keep them covered.
+  editable:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install scikit-build-core pybind11 numpy pytest
+      - run: pip install --no-build-isolation -ve .
+      - run: pytest tests -q
+
+  # Compile gate for the CUDA backend (no GPU on hosted runners, so tests run
+  # on the CPU paths; gpu_available() is False without a driver).
+  cuda-build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - uses: Jimver/cuda-toolkit@v0.2.35
+        with:
+          cuda: "12.9.1"
+          method: network
+          sub-packages: '["nvcc", "cudart"]'
+      - run: pip install scikit-build-core pybind11 numpy pytest
+      - run: pip install --no-build-isolation -v . -Ccmake.define.AUVUX_GPU=cuda
+      - run: pytest tests -q
+
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pipx run ruff check && pipx run ruff format --check
+      - run: pip install mypy numpy && mypy
+      - run: |
+          pip install clang-format
+          find src -name third_party -prune -o \( -name '*.cpp' -o -name '*.hpp' -o -name '*.cuh' -o -name '*.cu' -o -name '*.mm' -o -name '*.h' \) -print \
+            | xargs clang-format --dry-run --Werror

auvux_dsp-0.1.0.dev0/.github/workflows/wheels.yml

@@ -0,0 +1,47 @@
+name: wheels
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+
+# Build config lives in [tool.cibuildwheel] in pyproject.toml, so local
+# `cibuildwheel` runs and CI use the exact same settings.
+
+jobs:
+  wheels:
+    name: ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+    steps:
+      - uses: actions/checkout@v4
+      # CUDA toolkit for the Windows wheels (nvcc + cudart only; Linux gets
+      # CUDA from the manylinux_cuda images set in pyproject.toml). Build-time
+      # only — cudart is linked statically into the extension.
+      - uses: Jimver/cuda-toolkit@v0.2.35
+        if: runner.os == 'Windows'
+        with:
+          cuda: "12.9.1"
+          method: network
+          sub-packages: '["nvcc", "cudart"]'
+      - uses: pypa/cibuildwheel@v2.21.3
+      - uses: actions/upload-artifact@v4
+        with:
+          name: wheels-${{ matrix.os }}
+          path: wheelhouse/*.whl
+
+  sdist:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pipx run build --sdist
+      - uses: actions/upload-artifact@v4
+        with:
+          name: sdist
+          path: dist/*.tar.gz

auvux_dsp-0.1.0.dev0/.gitignore

@@ -0,0 +1,11 @@
+build/
+dist/
+wheelhouse/
+*.so
+*.pyd
+__pycache__/
+.pytest_cache/
+.venv/
+.mypy_cache/
+.ruff_cache/
+.DS_Store

auvux_dsp-0.1.0.dev0/CMakeLists.txt

@@ -0,0 +1,164 @@
+cmake_minimum_required(VERSION 3.18)
+project(auvux_dsp LANGUAGES C CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Release)
+endif()
+
+find_package(pybind11 CONFIG REQUIRED)
+find_package(Threads REQUIRED)
+
+# FFT: PFFFT is always compiled (vendored, guaranteed fallback); vDSP is added
+# on Apple. AUVUX_FFT=pffft drops vDSP even on Apple.
+set(AUVUX_FFT "auto" CACHE STRING "FFT backend: auto | vdsp | pffft")
+
+# GPU: Metal on Apple, CUDA where a toolkit is found, stub otherwise.
+# Override with -DAUVUX_GPU=metal|cuda|none; pick a CUDA toolkit with the
+# CUDACXX environment variable.
+set(AUVUX_GPU "auto" CACHE STRING "GPU backend: auto | metal | cuda | none")
+if(AUVUX_GPU STREQUAL "auto")
+  if(APPLE)
+    set(AUVUX_GPU "metal")
+  else()
+    include(CheckLanguage)
+    check_language(CUDA)
+    if(CMAKE_CUDA_COMPILER)
+      set(AUVUX_GPU "cuda")
+    else()
+      set(AUVUX_GPU "none")
+    endif()
+  endif()
+endif()
+message(STATUS "auvux-dsp GPU backend: ${AUVUX_GPU}")
+
+pybind11_add_module(_native
+  src/fft/fft.cpp
+  src/fft/fft_pffft.cpp
+  src/common/threadpool.cpp
+  src/common/host_pool.cpp
+  src/ops/window.cpp
+  src/ops/ola.cpp
+  src/ops/stft/stft_cpu.cpp
+  src/ops/istft/istft_cpu.cpp
+  src/ops/mel/mel_filterbank.cpp
+  src/ops/mel/mel_cpu.cpp
+  src/ops/cqt/cqt_filterbank.cpp
+  src/ops/cqt/cqt_cpu.cpp
+  src/ops/chroma/chroma_cpu.cpp
+  src/gpu/gpu_stub.cpp
+  src/gpu/cqt_plan.cpp
+  src/third_party/pffft.c
+  src/bindings/module.cpp
+  src/bindings/bind_fft.cpp
+  src/bindings/bind_stft.cpp
+  src/bindings/bind_mel.cpp
+  src/bindings/bind_cqt.cpp
+  src/bindings/bind_util.cpp)
+target_include_directories(_native PRIVATE src src/third_party)
+target_link_libraries(_native PRIVATE Threads::Threads)
+target_compile_definitions(_native PRIVATE AUVUX_HAVE_PFFFT=1)
+
+if(AUVUX_FFT STREQUAL "vdsp" AND NOT APPLE)
+  message(FATAL_ERROR "AUVUX_FFT=vdsp requires macOS")
+endif()
+if(AUVUX_FFT STREQUAL "vdsp" OR (AUVUX_FFT STREQUAL "auto" AND APPLE))
+  target_sources(_native PRIVATE src/fft/fft_vdsp.cpp)
+  target_compile_definitions(_native PRIVATE AUVUX_HAVE_VDSP=1)
+  find_library(ACCELERATE Accelerate REQUIRED)
+  target_link_libraries(_native PRIVATE ${ACCELERATE})
+  message(STATUS "auvux-dsp FFT backends: vdsp, pffft")
+elseif(AUVUX_FFT STREQUAL "auto" OR AUVUX_FFT STREQUAL "pffft")
+  message(STATUS "auvux-dsp FFT backends: pffft")
+else()
+  message(FATAL_ERROR "Unknown AUVUX_FFT='${AUVUX_FFT}' (use auto, vdsp, or pffft)")
+endif()
+
+if(AUVUX_GPU STREQUAL "metal")
+  if(NOT APPLE)
+    message(FATAL_ERROR "AUVUX_GPU=metal requires macOS")
+  endif()
+  enable_language(OBJCXX)
+  set(CMAKE_OBJCXX_STANDARD 17)
+  set(CMAKE_OBJCXX_STANDARD_REQUIRED ON)
+  # Kernel sources live in real .metal files; a generated header embeds them
+  # as strings for runtime compilation (no Metal toolchain needed at build).
+  set(auvux_metal_kernels
+      ${CMAKE_CURRENT_SOURCE_DIR}/src/gpu/metal/kernels/common.metal
+      ${CMAKE_CURRENT_SOURCE_DIR}/src/gpu/metal/kernels/stft.metal
+      ${CMAKE_CURRENT_SOURCE_DIR}/src/gpu/metal/kernels/mel.metal
+      ${CMAKE_CURRENT_SOURCE_DIR}/src/gpu/metal/kernels/cqt.metal)
+  set(auvux_metal_gen ${CMAKE_CURRENT_BINARY_DIR}/generated/auvux_metal_kernels.h)
+  string(JOIN "," auvux_metal_kernels_arg ${auvux_metal_kernels})
+  add_custom_command(
+    OUTPUT ${auvux_metal_gen}
+    COMMAND ${CMAKE_COMMAND} -DOUT=${auvux_metal_gen} "-DINPUTS=${auvux_metal_kernels_arg}"
+            -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/embed_text.cmake
+    DEPENDS ${auvux_metal_kernels} ${CMAKE_CURRENT_SOURCE_DIR}/cmake/embed_text.cmake
+    COMMENT "Embedding Metal kernel sources"
+    VERBATIM)
+  set(auvux_metal_srcs
+      src/gpu/metal/metal_common.mm
+      src/gpu/metal/metal_stft.mm
+      src/gpu/metal/metal_mel.mm
+      src/gpu/metal/metal_cqt.mm)
+  target_sources(_native PRIVATE ${auvux_metal_srcs} ${auvux_metal_gen})
+  set_source_files_properties(${auvux_metal_srcs} PROPERTIES COMPILE_OPTIONS "-fobjc-arc")
+  target_include_directories(_native PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/generated)
+  target_compile_definitions(_native PRIVATE AUVUX_METAL=1)
+  find_library(METAL_FRAMEWORK Metal REQUIRED)
+  find_library(FOUNDATION_FRAMEWORK Foundation REQUIRED)
+  target_link_libraries(_native PRIVATE ${METAL_FRAMEWORK} ${FOUNDATION_FRAMEWORK})
+elseif(AUVUX_GPU STREQUAL "cuda")
+  # Detect a user-supplied arch list before enable_language fills in the
+  # toolkit default and masks the distinction.
+  if(DEFINED CMAKE_CUDA_ARCHITECTURES AND CMAKE_CUDA_ARCHITECTURES)
+    set(auvux_user_archs TRUE)
+  else()
+    set(auvux_user_archs FALSE)
+  endif()
+  enable_language(CUDA)
+  target_sources(_native PRIVATE
+    src/gpu/cuda/cuda_common.cu
+    src/gpu/cuda/cuda_stft.cu
+    src/gpu/cuda/cuda_mel.cu
+    src/gpu/cuda/cuda_cqt.cu)
+  target_compile_definitions(_native PRIVATE AUVUX_CUDA=1)
+  if(NOT auvux_user_archs)
+    # Distributable fatbin: SASS per supported generation plus PTX (the plain
+    # "90" entry embeds both) so future architectures JIT.
+    set(auvux_archs 75-real 80-real 86-real 89-real 90)
+    set(AUVUX_MIN_CC 75)
+    if(CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 13)
+      list(PREPEND auvux_archs 60-real 70-real)  # CUDA 13 dropped pre-Turing
+      set(AUVUX_MIN_CC 60)
+    endif()
+    if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
+      list(APPEND auvux_archs 100-real 120-real)  # Blackwell
+    endif()
+    set(CMAKE_CUDA_ARCHITECTURES ${auvux_archs})
+    # Devices below the oldest shipped SASS get a clean CPU fallback.
+    target_compile_definitions(_native PRIVATE AUVUX_MIN_CC=${AUVUX_MIN_CC})
+  endif()
+  set_property(TARGET _native PROPERTY CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES})
+  # Static cudart: only the NVIDIA driver is needed at runtime; without one
+  # the wheel degrades to gpu_available() == False.
+  set_property(TARGET _native PROPERTY CUDA_RUNTIME_LIBRARY Static)
+elseif(NOT AUVUX_GPU STREQUAL "none")
+  message(FATAL_ERROR "Unknown AUVUX_GPU='${AUVUX_GPU}' (use auto, metal, cuda, or none)")
+endif()
+
+if(MSVC)
+  target_compile_options(_native PRIVATE
+    $<$<COMPILE_LANGUAGE:C,CXX>:/O2;/fp:fast>
+    $<$<COMPILE_LANGUAGE:CUDA>:-O3;-Xcompiler=/fp:fast>)
+else()
+  target_compile_options(_native PRIVATE
+    $<$<COMPILE_LANGUAGE:C,CXX>:-O3;-ffast-math;-funroll-loops>
+    $<$<COMPILE_LANGUAGE:CUDA>:-O3>)
+endif()
+
+install(TARGETS _native DESTINATION auvux/dsp)

auvux_dsp-0.1.0.dev0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Peter Kiers (Auvux)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

auvux_dsp-0.1.0.dev0/PKG-INFO

@@ -0,0 +1,90 @@
+Metadata-Version: 2.4
+Name: auvux-dsp
+Version: 0.1.0.dev0
+Summary: Fast differentiable audio transforms (STFT, mel, MFCC, CQT, chroma) on CPU and GPU
+Keywords: audio,dsp,stft,mel,mfcc,cqt,chroma,spectrogram,gpu
+Author-Email: Peter Kiers <pkiers.1983@gmail.com>
+License-Expression: MIT
+License-File: LICENSE
+License-File: THIRD_PARTY_LICENSES
+Classifier: Programming Language :: C++
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Multimedia :: Sound/Audio :: Analysis
+Requires-Python: >=3.10
+Requires-Dist: numpy>=1.22
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+Requires-Dist: librosa; extra == "test"
+Requires-Dist: soxr; extra == "test"
+Requires-Dist: torch; extra == "test"
+Description-Content-Type: text/markdown
+
+# auvux-dsp
+
+Fast differentiable audio transforms (STFT, iSTFT, mel, MFCC, CQT/VQT, chroma)
+with native CPU backends (vDSP on macOS, PFFFT elsewhere) and GPU support
+(Metal, CUDA). Forward *and* backward passes run in native kernels; PyTorch
+autograd plugs in when you pass a torch tensor, and torch is never required
+otherwise.
+
+```python
+import auvux.dsp as dsp
+
+mel = dsp.MelSpectrogram(sr=44100, n_fft=2048, hop_length=512, n_mels=128)
+S = mel(y)                       # numpy in -> numpy out
+S = mel(y, backend="gpu")        # Metal / CUDA kernels
+
+y = torch.tensor(clip, requires_grad=True)
+loss = mel(y, output="db").sum() # native forward
+loss.backward()                  # native adjoint kernel, no torch recompute
+```
+
+```
+pip install auvux-dsp --pre    # preview release; drop --pre once 0.1.0 is out
+```
+
+## Performance
+
+Benchmarks against librosa and torchaudio (forward passes and full training
+steps, CPU / staged GPU / GPU-resident) live in `benchmarks/benchmark.py`:
+
+```
+python benchmarks/benchmark.py
+```
+
+*Results table to be published with the first release.*
+
+### Complex STFT layout
+
+Complex spectra are returned as `(..., bins, frames)` backed by frame-major
+memory — each frame's spectrum contiguous, the freq axis strided. This is the
+same physical layout both references use (librosa allocates its stft output
+`order='F'`; `torch.stft` returns a transposed view over frame-major memory),
+so values *and* bytes match librosa, and the GPU-resident path returns a
+tensor with the exact strides `torch.stft` produces. It is also what makes
+the STFT fast: no backend ever materializes the bins-major transpose. `istft`
+accepts both this layout (zero-copy) and compact C-order arrays. Float
+outputs (power/db/mel/...) are ordinary C-contiguous arrays.
+
+Status: under construction.
+- CPU (vDSP/PFFFT): STFT, ISTFT, MelSpectrogram, MFCC, CQT, VQT, Chroma —
+  forward and native backward, librosa-parity tested, torch autograd built in.
+- Metal: all of the above on GPU (n_fft <= 4096), forward + backward,
+  parity-tested against the CPU path. torch MPS tensors stay on the GPU end to
+  end (DLPack), and backend="auto" routes them there — no flags needed.
+- CUDA: kernel-for-kernel twin of the Metal backend including the resident
+  paths; parity-tested on NVIDIA hardware (RTX 4090, CUDA 12.9), with pinned
+  double-buffered staging for the numpy-in/numpy-out GPU paths.
+- Pending: iCQT.
+
+## Development
+
+```
+pip install scikit-build-core pybind11 numpy pytest
+./scripts/dev-build.sh
+pytest
+```
+
+Note for packagers: `python/auvux/` is a PEP 420 namespace package — it must
+never contain an `__init__.py`, or it will shadow sibling `auvux-*`
+distributions.

auvux_dsp-0.1.0.dev0/README.md

@@ -0,0 +1,69 @@
+# auvux-dsp
+
+Fast differentiable audio transforms (STFT, iSTFT, mel, MFCC, CQT/VQT, chroma)
+with native CPU backends (vDSP on macOS, PFFFT elsewhere) and GPU support
+(Metal, CUDA). Forward *and* backward passes run in native kernels; PyTorch
+autograd plugs in when you pass a torch tensor, and torch is never required
+otherwise.
+
+```python
+import auvux.dsp as dsp
+
+mel = dsp.MelSpectrogram(sr=44100, n_fft=2048, hop_length=512, n_mels=128)
+S = mel(y)                       # numpy in -> numpy out
+S = mel(y, backend="gpu")        # Metal / CUDA kernels
+
+y = torch.tensor(clip, requires_grad=True)
+loss = mel(y, output="db").sum() # native forward
+loss.backward()                  # native adjoint kernel, no torch recompute
+```
+
+```
+pip install auvux-dsp --pre    # preview release; drop --pre once 0.1.0 is out
+```
+
+## Performance
+
+Benchmarks against librosa and torchaudio (forward passes and full training
+steps, CPU / staged GPU / GPU-resident) live in `benchmarks/benchmark.py`:
+
+```
+python benchmarks/benchmark.py
+```
+
+*Results table to be published with the first release.*
+
+### Complex STFT layout
+
+Complex spectra are returned as `(..., bins, frames)` backed by frame-major
+memory — each frame's spectrum contiguous, the freq axis strided. This is the
+same physical layout both references use (librosa allocates its stft output
+`order='F'`; `torch.stft` returns a transposed view over frame-major memory),
+so values *and* bytes match librosa, and the GPU-resident path returns a
+tensor with the exact strides `torch.stft` produces. It is also what makes
+the STFT fast: no backend ever materializes the bins-major transpose. `istft`
+accepts both this layout (zero-copy) and compact C-order arrays. Float
+outputs (power/db/mel/...) are ordinary C-contiguous arrays.
+
+Status: under construction.
+- CPU (vDSP/PFFFT): STFT, ISTFT, MelSpectrogram, MFCC, CQT, VQT, Chroma —
+  forward and native backward, librosa-parity tested, torch autograd built in.
+- Metal: all of the above on GPU (n_fft <= 4096), forward + backward,
+  parity-tested against the CPU path. torch MPS tensors stay on the GPU end to
+  end (DLPack), and backend="auto" routes them there — no flags needed.
+- CUDA: kernel-for-kernel twin of the Metal backend including the resident
+  paths; parity-tested on NVIDIA hardware (RTX 4090, CUDA 12.9), with pinned
+  double-buffered staging for the numpy-in/numpy-out GPU paths.
+- Pending: iCQT.
+
+## Development
+
+```
+pip install scikit-build-core pybind11 numpy pytest
+./scripts/dev-build.sh
+pytest
+```
+
+Note for packagers: `python/auvux/` is a PEP 420 namespace package — it must
+never contain an `__init__.py`, or it will shadow sibling `auvux-*`
+distributions.

auvux_dsp-0.1.0.dev0/THIRD_PARTY_LICENSES

@@ -0,0 +1,55 @@
+auvux-dsp bundles the following third-party components.
+
+================================================================================
+PFFFT — src/third_party/pffft.{c,h}
+Copyright (c) 2013 Julien Pommier
+Based on FFTPACK by Dr Paul Swarztrauber (NCAR), under the FFTPACK license.
+================================================================================
+
+Copyright (c) 2004 the University Corporation for Atmospheric Research
+("UCAR"). All rights reserved. Developed by NCAR's Computational and
+Information Systems Laboratory, UCAR, www.cisl.ucar.edu.
+
+Redistribution and use of the Software in source and binary forms, with or
+without modification, is permitted provided that the following conditions are
+met:
+
+- Neither the names of NCAR's Computational and Information Systems Laboratory,
+  the University Corporation for Atmospheric Research, nor the names of its
+  sponsors or contributors may be used to endorse or promote products derived
+  from this Software without specific prior written permission.
+- Redistributions of source code must retain the above copyright notices, this
+  list of conditions, and the disclaimer below.
+- Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions, and the disclaimer below in the documentation and/or
+  other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING, BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, INDIRECT,
+INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES OR OTHER LIABILITY,
+WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE SOFTWARE.
+
+================================================================================
+DLPack — src/third_party/dlpack.h
+Copyright (c) 2017 by DLPack Contributors
+================================================================================
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not
+use this file except in compliance with the License. You may obtain a copy of
+the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+License for the specific language governing permissions and limitations under
+the License.
+
+================================================================================
+pybind11 (build-time only; not redistributed in the wheel)
+Copyright (c) 2016 Wenzel Jakob — BSD-3-Clause.
+================================================================================