.flake8

 [flake8]
 max-line-length=120
-exclude=pystencils/jupyter.py,
-        pystencils/plot.py
-        pystencils/session.py
+exclude=src/pystencils/jupyter.py,
+        src/pystencils/plot.py
+        src/pystencils/session.py
 ignore = W293 W503 W291 C901 E741

.gitattributes

-pystencils/_version.py export-subst
+src/pystencils/_version.py export-subst

.gitignore

@@ -5,7 +5,7 @@ __pycache__
 *.vti
 /build
 /dist
-/*.egg-info
+*.egg-info
 .cache
 _build
 /html_doc
@@ -15,15 +15,16 @@ _build
 _local_tmp
 RELEASE-VERSION
 test-report
-pystencils/boundaries/createindexlistcython.c
-pystencils/boundaries/createindexlistcython.*.so
-pystencils_tests/tmp
-pystencils_tests/var
-pystencils_tests/kerncraft_inputs/.2d-5pt.c_kerncraft/
-pystencils_tests/kerncraft_inputs/.3d-7pt.c_kerncraft/
+src/pystencils/boundaries/createindexlistcython.c
+src/pystencils/boundaries/createindexlistcython.*.so
+tests/tmp
+tests/var
+tests/kerncraft_inputs/.2d-5pt.c_kerncraft/
+tests/kerncraft_inputs/.3d-7pt.c_kerncraft/
 report.xml
 coverage_report/
 
 
 # macOS
 **/.DS_Store
+*.uuid

.gitlab-ci.yml

 stages:
   - pretest
   - test
+  - nightly
+  - docs
   - deploy
 
 
+# --------------------------  Templates ------------------------------------------------------------------------------------
+
+# Base configuration for jobs meant to run at every commit
+.every-commit:
+  rules:
+    - if: $CI_PIPELINE_SOURCE != "schedule"
+
+# Configuration for jobs meant to run on each commit to pycodegen/pystencils/master
+.every-commit-master:
+  rules:
+    - if: '$CI_PIPELINE_SOURCE != "schedule" && $CI_PROJECT_PATH == "pycodegen/pystencils" && $CI_COMMIT_BRANCH == "master"'
+
+# Base configuration for jobs meant to run at a schedule
+.scheduled:
+  rules:
+    - if: $CI_PIPELINE_SOURCE == "schedule"
+
 # --------------------------  Tests ------------------------------------------------------------------------------------
 
 # Normal test - runs on every commit all but "long run" tests
 tests-and-coverage:
   stage: pretest
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
-  image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full
+  extends: .every-commit
+  image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full:cupy12.3
+  before_script:
+    - pip install -e .
   script:
-    - pip install sympy --upgrade
     - env
     - pip list
     - export NUM_CORES=$(nproc --all)
     - mkdir -p ~/.config/matplotlib
     - echo "backend:template" > ~/.config/matplotlib/matplotlibrc
     - mkdir public
-    - py.test -v -n $NUM_CORES --cov-report html --cov-report term --cov=. -m "not longrun" --html test-report/index.html --junitxml=report.xml
-    - python3 -m coverage xml
+    - pytest -v -n $NUM_CORES --cov-report html --cov-report xml --cov-report term --cov=. -m "not longrun" --html test-report/index.html --junitxml=report.xml
+    - python -m coverage xml
   tags:
     - docker
     - cuda11
@@ -44,9 +62,11 @@ tests-and-coverage-with-longrun:
   stage: test
   when: manual
   allow_failure: true
-  image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full
-  script:
+  image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full:cupy12.3
+  before_script:
     - pip install sympy --upgrade
+    - pip install -e .
+  script:
     - env
     - pip list
     - export NUM_CORES=$(nproc --all)
@@ -62,13 +82,14 @@ tests-and-coverage-with-longrun:
 # pipeline with latest python version
 latest-python:
   stage: test
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/latest_python
+  before_script:
+    - pip install -e .
   script:
     - env
     - pip list
+    - pip install -e .
     - export NUM_CORES=$(nproc --all)
     - mkdir -p ~/.config/matplotlib
     - echo "backend:template" > ~/.config/matplotlib/matplotlibrc
@@ -84,41 +105,34 @@ latest-python:
 
 
 # Minimal tests in windows environment
-minimal-windows:
-  stage: test
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
-  tags:
-    - win
-  script:
-    - export NUM_CORES=$(nproc --all)
-    - source /cygdrive/c/Users/build/Miniconda3/Scripts/activate
-    - source activate pystencils
-    - pip install joblib
-    - pip list
-    - python -c "import numpy"
-    - py.test -v -m "not (notebook or longrun)"
+#minimal-windows:
+#  stage: test
+#  tags:
+#    - win
+#  script:
+#    - export NUM_CORES=$(nproc --all)
+#    - source /cygdrive/c/Users/build/Miniconda3/Scripts/activate
+#    - source activate pystencils
+#    - pip install joblib
+#    - pip list
+#    - python -c "import numpy"
+#    - py.test -v -m "not (notebook or longrun)"
 
 ubuntu:
   stage: test
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/ubuntu
   before_script:
-    - apt-get -y remove python3-sympy
     - ln -s /usr/include/locale.h /usr/include/xlocale.h
-    - pip3 install `grep -Eo 'sympy[>=]+[0-9\.]+' setup.py | sed 's/>/=/g'`
-    # - pip3 install `grep -Eo 'sympy[>=]+[0-9\.]+' setup.py | sed 's/>/=/g'`
+    - pip3 install -e .
   script:
     - export NUM_CORES=$(nproc --all)
     - mkdir -p ~/.config/matplotlib
     - echo "backend:template" > ~/.config/matplotlib/matplotlibrc
     - sed -i 's/--doctest-modules //g' pytest.ini
     - env
-    - pip3 list
-    - pytest-3 -v -n $NUM_CORES -m "not longrun" --junitxml=report.xml
+    - pip list
+    - pytest -v -n $NUM_CORES -m "not longrun" --junitxml=report.xml
   tags:
     - docker
     - cuda11
@@ -130,10 +144,11 @@ ubuntu:
 
 .multiarch_template:
   stage: test
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
+  allow_failure: true
   before_script: &multiarch_before_script
+    # - pip3 install -v .
+    - export PYTHONPATH=src
     - python3 -c "import pystencils as ps; ps.cpu.cpujit.read_config()"
     - sed -i '/^fail_under.*/d' pytest.ini
   script:
@@ -143,20 +158,26 @@ ubuntu:
     - sed -i 's/--doctest-modules //g' pytest.ini
     - env
     - pip3 list
-    - pytest-3 -v -n $NUM_CORES --junitxml=report.xml pystencils_tests/test_*vec*.py pystencils_tests/test_random.py
+    - python3 -m pytest -v -n $NUM_CORES --cov-report html --cov-report xml --cov=. --junitxml=report.xml tests/test_*vec*.py tests/test_random.py tests/test_half_precision.py
+    - python3 -m coverage xml
   tags:
     - docker
     - AVX
   artifacts:
     when: always
+    paths:
+      - coverage_report
     reports:
+      coverage_report:
+        coverage_format: cobertura
+        path: coverage.xml
       junit: report.xml
 
 arm64v8:
   extends: .multiarch_template
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/arm64
   variables:
-    PYSTENCILS_SIMD: "neon"
+    QEMU_CPU: "cortex-a76"
   before_script:
     - *multiarch_before_script
     - sed -i s/march=native/march=armv8-a/g ~/.config/pystencils/config.json
@@ -164,45 +185,41 @@ arm64v8:
 ppc64le:
   extends: .multiarch_template
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/ppc64le
-  variables:
-    PYSTENCILS_SIMD: "vsx"
   before_script:
     - *multiarch_before_script
     - sed -i s/mcpu=native/mcpu=power8/g ~/.config/pystencils/config.json
 
 arm64v9:
   # SVE support is still unreliable in GCC 11 (incorrect code for fixed-width vectors, internal compiler errors).
+  # For half precision Clang is necessary
   extends: .multiarch_template
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/arm64
-  variables:
-    PYSTENCILS_SIMD: "sve128,sve256,sve512,sve"
   before_script:
     - *multiarch_before_script
-    - sed -i s/march=native/march=armv8-a+sve/g ~/.config/pystencils/config.json
+    - sed -i s/march=native/march=armv9-a+sve2+sme/g ~/.config/pystencils/config.json
     - sed -i s/g\+\+/clang++/g ~/.config/pystencils/config.json
 
 riscv64:
   # RISC-V vector extension are currently not supported by GCC.
-  # Also, the image is built without the libomp package which is not yet available on Ubuntu.
   extends: .multiarch_template
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/riscv64
   variables:
+    # explicitly set SIMD as detection requires QEMU >= 8.1
     PYSTENCILS_SIMD: "rvv"
-    QEMU_CPU: "rv64,v=true"
+    QEMU_CPU: "rv64,v=true,zicboz=true"
   before_script:
     - *multiarch_before_script
-    - sed -i 's/march=native/march=rv64imfdv/g' ~/.config/pystencils/config.json
-    - sed -i s/g\+\+/clang++/g ~/.config/pystencils/config.json
-    - sed -i 's/fopenmp/fopenmp=libgomp -I\/usr\/include\/riscv64-linux-gnu/g' ~/.config/pystencils/config.json
+    - sed -i 's/march=native/march=rv64imfdvzicboz/g' ~/.config/pystencils/config.json
+    - sed -i s/g\+\+/clang++-15/g ~/.config/pystencils/config.json
 
 minimal-conda:
   stage: pretest
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/minimal_conda
+  before_script:
+    - pip install -e .
   script:
-    - python setup.py quicktest
+    - python quicktest.py
   tags:
     - docker
     - cuda
@@ -210,13 +227,13 @@ minimal-conda:
 
 minimal-sympy-master:
   stage: test
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/minimal_conda
+  before_script:
+    - pip install -e .
   script:
     - python -m pip install --upgrade git+https://github.com/sympy/sympy.git
-    - python setup.py quicktest
+    - python quicktest.py
   allow_failure: true
   tags:
     - docker
@@ -267,26 +284,57 @@ pycodegen-integration:
     reports:
       junit: pycodegen/*/report.xml
 
+
+# -------------------- Scheduled Tasks --------------------------------------------------------------------------
+
+
+# Nightly test against the latest (pre-release) version of SymPy published on PyPI
+nightly-sympy:
+  stage: nightly
+  needs: []
+  extends: .scheduled
+  image: i10git.cs.fau.de:5005/pycodegen/pycodegen/latest_python
+  before_script:
+    - pip install -e .
+    - pip install --upgrade --pre sympy
+  script:
+    - env
+    - pip list
+    - export NUM_CORES=$(nproc --all)
+    - mkdir -p ~/.config/matplotlib
+    - echo "backend:template" > ~/.config/matplotlib/matplotlibrc
+    - mkdir public
+    - pytest -v -n $NUM_CORES -m "not longrun" --junitxml=report.xml
+  tags:
+    - docker
+    - AVX
+    - cuda
+  artifacts:
+    when: always
+    reports:
+      junit: report.xml
+
 # -------------------- Linter & Documentation --------------------------------------------------------------------------
 
 
 flake8-lint:
   stage: pretest
-  except:
-    variables:
-      - $ENABLE_NIGHTLY_BUILDS
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full
   script:
-    - flake8 pystencils
+    - flake8 src/pystencils
   tags:
     - docker
 
 
 build-documentation:
-  stage: test
+  stage: docs
+  extends: .every-commit
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/documentation
+  needs: []
+  before_script:
+    - pip install -e .
   script:
-    - export PYTHONPATH=`pwd`
     - mkdir html_doc
     - sphinx-build -b html doc  html_doc
     - sphinx-build -W -b html doc  html_doc
@@ -299,7 +347,9 @@ build-documentation:
 
 pages:
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/full
+  extends: .every-commit-master
   stage: deploy
+  needs: ["tests-and-coverage", "build-documentation"]
   script:
     - ls -l
     - mv coverage_report html_doc
@@ -309,5 +359,3 @@ pages:
       - public
   tags:
     - docker
-  only:
-    - master@pycodegen/pystencils

MANIFEST.in

-include README.md
-include COPYING.txt
 include AUTHORS.txt
 include CONTRIBUTING.md
-CHANGELOG.md
-global-include *.pyx
-include versioneer.py
-include pystencils/_version.py
+include CHANGELOG.md

README.md

@@ -52,7 +52,7 @@ pip install pystencils[interactive]
 Without `[interactive]` you get a minimal version with very little dependencies.
 
 All options:
-- `gpu`: use this if an NVIDIA GPU is available and CUDA is installed
+- `gpu`: use this if an NVIDIA or AMD GPU is available and CUDA or ROCm is installed
 - `alltrafos`: pulls in additional dependencies for loop simplification e.g. libisl
 - `bench_db`: functionality to store benchmark result in object databases
 - `interactive`: installs dependencies to work in Jupyter including image I/O, plotting etc.
@@ -63,7 +63,7 @@ Options can be combined e.g.
 pip install pystencils[interactive, gpu, doc]
 ```
 
-pystencils is also fully compatible with Windows machines. If working with visual studio and pycuda makes sure to run example files first to ensure that pycuda can find the compiler's executable.
+pystencils is also fully compatible with Windows machines. If working with visual studio and cupy makes sure to run example files first to ensure that cupy can find the compiler's executable.
 
 Documentation
 -------------
@@ -81,4 +81,7 @@ Many thanks go to the [contributors](https://i10git.cs.fau.de/pycodegen/pystenci
 If you use pystencils in a publication, please cite the following articles:
 
 Overview:
-  - M. Bauer et al, Code Generation for Massively Parallel Phase-Field Simulations. Association for Computing Machinery, 2019. https://doi.org/10.1145/3295500.3356186
+  - M. Bauer et al, Code Generation for Massively Parallel Phase-Field Simulations. Association for Computing Machinery, 2019. https://doi.org/10.1145/3295500.3356186
\ No newline at end of file
+
+Performance Modelling:
+  - D. Ernst et al, Analytical performance estimation during code generation on modern GPUs. Journal of Parallel and Distributed Computing, 2023. https://doi.org/10.1016/j.jpdc.2022.11.003

binder/environment.yml

@@ -7,7 +7,7 @@
 #     conda env create -f conda_environment_user.yml
 #     . activate pystencils
 #
-# If you have CUDA installed and want to use your GPU, uncomment the last line to install pycuda
+# If you have CUDA or ROCm installed and want to use your GPU, uncomment the last line to install cupy
 #
 # ----------------------------------------------------------------------------------------------------------------------
 
@@ -32,4 +32,4 @@ dependencies:
       - ipy_table  # HTML tables for jupyter notebooks
       - pyevtk # VTK output for serial simulations
       - blitzdb # file-based No-SQL database to store simulation results
-      #- pycuda # add this if you have CUDA installed
+      #- cupy # add this if you have CUDA or ROCm installed

conftest.py

@@ -8,7 +8,7 @@ import nbformat
 import pytest
 from nbconvert import PythonExporter
 
-from pystencils.boundaries.createindexlistcython import *  # NOQA
+from pystencils.boundaries.createindexlist import *  # NOQA
 # Trigger config file reading / creation once - to avoid race conditions when multiple instances are creating it
 # at the same time
 from pystencils.cpu import cpujit
@@ -40,47 +40,48 @@ def add_path_to_ignore(path):
 
 
 collect_ignore = [os.path.join(SCRIPT_FOLDER, "doc", "conf.py"),
-                  os.path.join(SCRIPT_FOLDER, "pystencils", "opencl", "opencl.autoinit")]
-add_path_to_ignore('pystencils_tests/benchmark')
+                  os.path.join(SCRIPT_FOLDER, "src", "pystencils", "opencl", "opencl.autoinit")]
+add_path_to_ignore('tests/benchmark')
 add_path_to_ignore('_local_tmp')
 
 
 try:
-    import pycuda
+    import cupy
 except ImportError:
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils_tests/test_cudagpu.py")]
-    add_path_to_ignore('pystencils/gpucuda')
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "tests/test_gpu.py")]
+    add_path_to_ignore('src/pystencils/gpu')
 
 try:
     import waLBerla
 except ImportError:
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils_tests/test_aligned_array.py"),
-                       os.path.join(SCRIPT_FOLDER, "pystencils_tests/test_datahandling_parallel.py"),
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "tests/test_aligned_array.py"),
+                       os.path.join(SCRIPT_FOLDER, "tests/test_datahandling_parallel.py"),
                        os.path.join(SCRIPT_FOLDER, "doc/notebooks/03_tutorial_datahandling.ipynb"),
-                       os.path.join(SCRIPT_FOLDER, "pystencils/datahandling/parallel_datahandling.py"),
-                       os.path.join(SCRIPT_FOLDER, "pystencils_tests/test_small_block_benchmark.ipynb")]
+                       os.path.join(SCRIPT_FOLDER, "src/pystencils/datahandling/parallel_datahandling.py"),
+                       os.path.join(SCRIPT_FOLDER, "tests/test_small_block_benchmark.ipynb")]
 
 try:
     import blitzdb
 except ImportError:
-    add_path_to_ignore('pystencils/runhelper')
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils_tests/test_parameterstudy.py")]
+    add_path_to_ignore('src/pystencils/runhelper')
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "tests/test_parameterstudy.py")]
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "tests/test_json_serializer.py")]
 
 try:
     import islpy
 except ImportError:
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils/integer_set_analysis.py")]
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "src/pystencils/integer_set_analysis.py")]
 
 try:
     import graphviz
 except ImportError:
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils/backends/dot.py")]
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "src/pystencils/backends/dot.py")]
     collect_ignore += [os.path.join(SCRIPT_FOLDER, "doc/notebooks/01_tutorial_getting_started.ipynb")]
 
 try:
     import pyevtk
 except ImportError:
-    collect_ignore += [os.path.join(SCRIPT_FOLDER, "pystencils/datahandling/vtk.py")]
+    collect_ignore += [os.path.join(SCRIPT_FOLDER, "src/pystencils/datahandling/vtk.py")]
 
 collect_ignore += [os.path.join(SCRIPT_FOLDER, 'setup.py')]
 

doc/conf.py

@@ -33,7 +33,7 @@ version = re.sub(r'(\d+\.\d+)\.\d+(.*)', r'\1\2', pystencils.__version__)
 version = re.sub(r'(\.dev\d+).*?$', r'\1', version)
 # The full version, including alpha/beta/rc tags.
 release = pystencils.__version__
-language = None
+language = 'en'
 exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store', '**.ipynb_checkpoints']
 default_role = 'any'
 pygments_style = 'sphinx'

doc/notebooks/02_tutorial_basic_kernels.ipynb

 %% Cell type:code id: tags:
 
 ``` python
 from pystencils.session import *
 ```
 
 %% Cell type:markdown id: tags:
 
 # Tutorial 02: Basic Kernel generation with *pystencils*
 
 Now that you have an [overview of pystencils](01_tutorial_getting_started.ipynb),
 this tutorial shows in more detail how to formulate, optimize and run stencil kernels.
 
 ## 1) Kernel Definition
 
 ### a) Defining kernels with assignment lists and the  `kernel` decorator
 
 *pystencils* gets a symbolic formulation of the kernel. This can be either an `Assignment` or a sequence of `Assignment`s that follow a set of restrictions.
 
 Lets first create a kernel that consists of multiple assignments:
 
 %% Cell type:code id: tags:
 
 ``` python
 src_arr = np.zeros([20, 30])
 dst_arr = np.zeros_like(src_arr)
 
 dst, src = ps.fields(dst=dst_arr, src=src_arr)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 grad_x, grad_y = sp.symbols("grad_x, grad_y")
 
 symbolic_description = [
     ps.Assignment(grad_x, (src[1, 0] - src[-1, 0]) / 2),
     ps.Assignment(grad_y, (src[0, 1] - src[0, -1]) / 2),
     ps.Assignment(dst[0, 0], grad_x + grad_y),
 ]
 kernel = ps.create_kernel(symbolic_description)
 symbolic_description
 ```
 
 %% Output
 
-
-    $\displaystyle \left[ grad_{x} \leftarrow \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2}, \  grad_{y} \leftarrow \frac{{src}_{(0,1)}}{2} - \frac{{src}_{(0,-1)}}{2}, \  {dst}_{(0,0)} \leftarrow grad_{x} + grad_{y}\right]$
+
+    $\displaystyle \left[ grad_{x} \leftarrow_{} \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2}, \  grad_{y} \leftarrow_{} \frac{{src}_{(0,1)}}{2} - \frac{{src}_{(0,-1)}}{2}, \  {dst}_{(0,0)} \leftarrow_{} grad_{x} + grad_{y}\right]$
     ⎡         src_E   src_W            src_N   src_S                         ⎤
     ⎢gradₓ := ───── - ─────, grad_y := ───── - ─────, dst_C := gradₓ + grad_y⎥
     ⎣           2       2                2       2                           ⎦
 
 %% Cell type:markdown id: tags:
 
 We created subexpressions, using standard sympy symbols on the left hand side, to split the kernel into multiple assignments. Defining a kernel using a list of `Assignment`s is quite tedious and hard to read.
 To simplify the formulation of a kernel, *pystencils* offers the `kernel` decorator, that transforms a normal Python function with `@=` assignments into an assignment list that can be passed to `create_kernel`.
 
 %% Cell type:code id: tags:
 
 ``` python
 @ps.kernel
 def symbolic_description_using_function():
     grad_x @= (src[1, 0] - src[-1, 0]) / 2
     grad_y @= (src[0, 1] - src[0, -1]) / 2
     dst[0, 0] @= grad_x + grad_y
 symbolic_description_using_function
 ```
 
 %% Output
 
-
-    $\displaystyle \left[ grad_{x} \leftarrow \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2}, \  grad_{y} \leftarrow \frac{{src}_{(0,1)}}{2} - \frac{{src}_{(0,-1)}}{2}, \  {dst}_{(0,0)} \leftarrow grad_{x} + grad_{y}\right]$
+
+    $\displaystyle \left[ grad_{x} \leftarrow_{} \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2}, \  grad_{y} \leftarrow_{} \frac{{src}_{(0,1)}}{2} - \frac{{src}_{(0,-1)}}{2}, \  {dst}_{(0,0)} \leftarrow_{} grad_{x} + grad_{y}\right]$
     ⎡         src_E   src_W            src_N   src_S                         ⎤
     ⎢gradₓ := ───── - ─────, grad_y := ───── - ─────, dst_C := gradₓ + grad_y⎥
     ⎣           2       2                2       2                           ⎦
 
 %% Cell type:markdown id: tags:
 
 The decorated function can contain any Python code, only the `@=` operator, and the ternary inline `if-else` operator have different meaning.
 
 ### b) Ternary 'if' with `Piecewise`
 
 The ternary operator maps to `sympy.Piecewise` functions, that can be used to introduce branching into the kernel. Piecewise defined functions must give a value for every input, i.e. there must be a 'otherwise' clause in the end that is indicated by the condition `True`. Piecewise objects are standard sympy terms that can be integrated into bigger expressions:
 
 %% Cell type:code id: tags:
 
 ``` python
 sp.Piecewise((1.0, src[0,1] > 0), (0.0, True)) + src[1, 0]
 ```
 
 %% Output
 
     $\displaystyle {src}_{(1,0)} + \begin{cases} 1.0 & \text{for}\: {src}_{(0,1)} > 0 \\0.0 & \text{otherwise} \end{cases}$
             ⎛⎧1.0  for src_N > 0⎞
     src_E + ⎜⎨                  ⎟
             ⎝⎩0.0    otherwise  ⎠
 
 %% Cell type:markdown id: tags:
 
 Piecewise objects are created by the `kernel` decorator for ternary if-else statements.
 
 %% Cell type:code id: tags:
 
 ``` python
 @ps.kernel
 def kernel_with_piecewise():
     grad_x @= (src[1, 0] - src[-1, 0]) / 2 if src[-1, 0] > 0 else 0.0
 kernel_with_piecewise
 ```
 
 %% Output
 
-
-    $\displaystyle \left[ grad_{x} \leftarrow \begin{cases} \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2} & \text{for}\: {src}_{(-1,0)} > 0 \\0.0 & \text{otherwise} \end{cases}\right]$
+
+    $\displaystyle \left[ grad_{x} \leftarrow_{} \begin{cases} \frac{{src}_{(1,0)}}{2} - \frac{{src}_{(-1,0)}}{2} & \text{for}\: {src}_{(-1,0)} > 0 \\0.0 & \text{otherwise} \end{cases}\right]$
     ⎡         ⎧src_E   src_W               ⎤
     ⎢         ⎪───── - ─────  for src_W > 0⎥
     ⎢gradₓ := ⎨  2       2                 ⎥
     ⎢         ⎪                            ⎥
     ⎣         ⎩     0.0         otherwise  ⎦
 
 %% Cell type:markdown id: tags:
 
 ### c) Assignment level optimizations using `AssignmentCollection`
 
 When the kernels get larger and more complex, it is helpful to organize the list of assignment into a more structured way. The `AssignmentCollection` offers optimizating transformation on a list of assignments. It holds two assignment lists, one for subexpressions and one for the main assignments. Main assignments are typically those that write to an array.
 
 %% Cell type:code id: tags:
 
 ``` python
 @ps.kernel
 def somewhat_longer_dummy_kernel(s):
     s.a @= src[0, 1] + src[-1, 0]
     s.b @= 2 * src[1, 0] + src[0, -1]
     s.c @= src[0, 1] + 2 * src[1, 0] + src[-1, 0] + src[0, -1] - src[0,0]
     dst[0, 0] @= s.a + s.b + s.c
 
 ac = ps.AssignmentCollection(main_assignments=somewhat_longer_dummy_kernel[-1:],
                              subexpressions=somewhat_longer_dummy_kernel[:-1])
 ac
 ```
 
 %% Output
 
-    AssignmentCollection: dst_C, <- f(src_N, src_E, src_W, src_C, src_S)
+    AssignmentCollection: dst_C, <- f(src_C, src_W, src_S, src_N, src_E)
 
 %% Cell type:code id: tags:
 
 ``` python
 ac.operation_count
 ```
 
 %% Output
 
     {'adds': 8,
      'muls': 2,
      'divs': 0,
      'sqrts': 0,
      'fast_sqrts': 0,
      'fast_inv_sqrts': 0,
      'fast_div': 0}
 
 %% Cell type:markdown id: tags:
 
 The `pystencils.simp` submodule offers several functions to optimize a collection of assignments.
 It also offers functionality to group optimization into strategies and evaluate them.
 In this example we reduce the number of operations by reusing existing subexpressions to get rid of two unnecessary floating point additions. For more information about assignment collections and simplifications see the [demo notebook](demo_assignment_collection.ipynb).
 
 %% Cell type:code id: tags:
 
 ``` python
 opt_ac = ps.simp.subexpression_substitution_in_existing_subexpressions(ac)
 opt_ac
 ```
 
 %% Output
 
-    AssignmentCollection: dst_C, <- f(src_N, src_E, src_W, src_C, src_S)
+    AssignmentCollection: dst_C, <- f(src_C, src_W, src_S, src_N, src_E)
 
 %% Cell type:code id: tags:
 
 ``` python
 opt_ac.operation_count
 ```
 
 %% Output
 
     {'adds': 6,
      'muls': 1,
      'divs': 0,
      'sqrts': 0,
      'fast_sqrts': 0,
      'fast_inv_sqrts': 0,
      'fast_div': 0}
 
 %% Cell type:markdown id: tags:
 
 ### d) Ghost layers and iteration region
 
 When creating a kernel with neighbor accesses, *pystencils* automatically restricts the iteration region, such that all accesses are safe.
 
 %% Cell type:code id: tags:
 
 ``` python
 kernel = ps.create_kernel(ps.Assignment(dst[0,0], src[2, 0] + src[-1, 0]))
 ps.show_code(kernel)
 ```
 
 %% Output
 
 
 
 %% Cell type:markdown id: tags:
 
 When no additional ghost layer information is given, *pystencils* looks at all neighboring field accesses and introduces the required number of ghost layers **for all directions**. In the example above the largest neighbor accesses was ``src[2, 0]``, so theoretically we would need 2 ghost layers only the the end of the x coordinate.
 By default *pystencils* introduces 2 ghost layers at all borders of the domain. The next cell shows how to change this behavior. Be careful with manual ghost layer specification, wrong values may lead to SEGFAULTs.
 
 %% Cell type:code id: tags:
 
 ``` python
 gl_spec = [(0, 2),   # 0 ghost layers at the left, 2 at the right border
            (1, 0)]   # 1 ghost layer at the lower y, one at the upper y coordinate
 kernel = ps.create_kernel(ps.Assignment(dst[0,0], src[2, 0] + src[-1, 0]), ghost_layers=gl_spec)
 ps.show_code(kernel)
 ```
 
 %% Output
 
 
 
 %% Cell type:markdown id: tags:
 
 ## 2 ) Restrictions
 
 
 ### a) Independence Restriction
 
 *pystencils* only works for kernels where each array element can be updated independently from all other elements. This restriction ensures that the kernels can be easily parallelized and also be run on the GPU. Trying to define kernels where the results depends on the iteration order, leads to a ValueError.
 
 %% Cell type:code id: tags:
 
 ``` python
 invalid_description = [
     ps.Assignment(dst[1, 0], src[1, 0] + src[-1, 0]),
     ps.Assignment(dst[0, 0], src[1, 0] - src[-1, 0]),
 ]
 try:
     invalid_kernel = ps.create_kernel(invalid_description)
     assert False, "Should never be executed"
 except ValueError as e:
     print(e)
 ```
 
 %% Output
 
     Field dst is written at two different locations
 
 %% Cell type:markdown id: tags:
 
 The independence restriction makes sure that the kernel can be safely parallelized by checking the following conditions: If a field is modified inside the kernel, it may only be modified at a single spatial position. In that case the field may also only be read at this position. Fields that are not modified may be read at multiple neighboring positions.
 
 Specifically, this rule allows for in-place updates that don't access neighbors.
 
 %% Cell type:code id: tags:
 
 ``` python
 valid_kernel = ps.create_kernel(ps.Assignment(src[0,0], 2*src[0,0] + 42))
 ```
 
 %% Cell type:markdown id: tags:
 
 If a field stores multiple values per cell, as in the next example, this restriction only applies for accesses with the same index.
 
 %% Cell type:code id: tags:
 
 ``` python
 v = ps.fields("v(2): double[2D]")
 valid_kernel = ps.create_kernel([ps.Assignment(v[0,0](1), 2*v[0,0](1) + 42),
-                                 ps.Assignment(v[0,1](0), 2*v[1,0](0) + 42)])
+                                 ps.Assignment(v[0,1](0), 2*v[0,1](0) + 42)])
 ```
 
 %% Cell type:markdown id: tags:
 
 ### b) Static Single Assignment Form
 
 All assignments that don't write to a field must be in SSA form
 1. Each sympy symbol may only occur once as a left-hand-side (fields can be written multiple times)
 2. A symbol has to be defined before it is used. If it is never defined it is introduced as function parameter
 
 The next cell demonstrates the first SSA restriction:
 
 %% Cell type:code id: tags:
 
 ``` python
 @ps.kernel
 def not_allowed():
     a, b = sp.symbols("a b")
     a @= src[0, 0]
     b @= a + 3
     a @= src[-1, 0]
     dst[0, 0] @= a + b
 try:
     ps.create_kernel(not_allowed)
     assert False
 except ValueError as e:
     print(e)
 ```
 
 %% Output
 
     Assignments not in SSA form, multiple assignments to a
 
 %% Cell type:markdown id: tags:
 
 Also it is not allowed to write a field at the same location
 
 %% Cell type:code id: tags:
 
 ``` python
 @ps.kernel
 def not_allowed():
     dst[0, 0] @= src[0, 1] + src[1, 0]
     dst[0, 0] @= 2 * dst[0, 0]
 
 try:
     ps.create_kernel(not_allowed)
     assert False
 except ValueError as e:
     print(e)
 ```
 
 %% Output
 
     Field dst is written twice at the same location
 
 %% Cell type:markdown id: tags:
 
 This situation should be resolved by introducing temporary variables
 
 %% Cell type:code id: tags:
 
 ``` python
 tmp_var = sp.Symbol("a")
 
 @ps.kernel
 def allowed():
     tmp_var @= src[0, 1] + src[1, 0]
     dst[0, 0] @= 2 * tmp_var
 
 
 ast = ps.create_kernel(allowed)
 ps.show_code(ast)
 ```
 
 %% Output
 
 

doc/notebooks/03_tutorial_datahandling.ipynb

 %% Cell type:code id: tags:
 
 ``` python
 import psutil
 from mpl_toolkits.mplot3d import Axes3D
 from matplotlib import pyplot, cm
 from pystencils.session import *
 from pystencils.boundaries import add_neumann_boundary, Neumann, Dirichlet, BoundaryHandling
 from pystencils.slicing import slice_from_direction
 import math
 import time
 %matplotlib inline
 ```
 
 %% Cell type:markdown id: tags:
 
-Test to see if pycuda is installed which is needed to run calculations on the GPU
+Test to see if cupy is installed which is needed to run calculations on the GPU
 
 %% Cell type:code id: tags:
 
 ``` python
 try:
-    import pycuda
+    import cupy
     gpu = True
 except ImportError:
     gpu = False
-    pycuda = None
-    print('No pycuda installed')
-
-if pycuda:
-    import pycuda.gpuarray as gpuarray
+    cupy = None
+    print('No cupy installed')
 ```
 
-%% Output
-
-    No pycuda installed
-
 %% Cell type:markdown id: tags:
 
 # Tutorial 03: Datahandling
 
 %% Cell type:markdown id: tags:
 
 This is a tutorial about the `DataHandling` class of pystencils. This class is an abstraction layer to
 - link numpy arrays to pystencils fields
 - handle CPU-GPU array transfer, such that one can write code that works on CPU and GPU
 - makes it possible to write MPI parallel simulations to run on distributed-memory clusters using the waLBerla library
 
 We will look at a small and easy example to demonstrate the usage of `DataHandling` objects. We will define an averaging kernel to every cell of an array, that writes the average of the neighbor cell values to the center.
 
 %% Cell type:markdown id: tags:
 
 ## 1. Manual
 
 ### 1.1. CPU kernels
 
 In this first part, we set up a scenario manually without a `DataHandling`. In the next sections we then repeat the same setup with the help of the data handling.
 
 One concept of *pystencils* that may be confusing at first, is the differences between pystencils fields and numpy arrays. Fields are used to describe the computation *symbolically* with sympy, while numpy arrays hold the actual values where the computation is executed on.
 
 One option to create and execute a *pystencils* kernel is listed below. For reasons that become clear later we call this the **variable-field-size workflow**:
 
 1. define pystencils fields
 2. use sympy and the pystencils fields to define an update rule, that describes what should be done on *every cell*
 3. compile the update rule to a real function, that can be called from Python. For each field that was referenced in the symbolic description the function expects a numpy array, passed as named parameter
 4. create some numpy arrays with actual data
 5. call the kernel - usually many times
 
 Now, lets see how this actually looks in Python code:
 
 %% Cell type:code id: tags:
 
 ``` python
 # 1. field definitions
 src_field, dst_field = ps.fields("src, dst:[2D]")
 
 # 2. define update rule
 update_rule = [ps.Assignment(lhs=dst_field[0, 0],
                              rhs=(src_field[1, 0] + src_field[-1, 0] +
                                   src_field[0, 1] + src_field[0, -1]) / 4)]
 
 # 3. compile update rule to function
 kernel_function = ps.create_kernel(update_rule).compile()
 
 # 4. create numpy arrays and call kernel
 src_arr, dst_arr = np.random.rand(30, 30), np.zeros([30, 30])
 
 # 5. call kernel
 kernel_function(src=src_arr, dst=dst_arr)  # names of arguments have to match names passed to ps.fields()
 ```
 
 %% Cell type:markdown id: tags:
 
 This workflow separates the symbolic and the numeric stages very cleanly. The separation also makes it possible to stop after step 3, write the C-code to a file and call the kernel from a C program. Speaking of the C-Code - lets have a look at the generated sources:
 
 %% Cell type:code id: tags:
 
 ``` python
 ps.show_code(kernel_function.ast)
 ```
 
 %% Output
 
 
 
 %% Cell type:markdown id: tags:
 
 Even if it looks very ugly and low-level :) lets look at this code in a bit more detail. The code is generated in a way that it works for different array sizes. The size of the array is passed in the `_size_dst_` variables that specifiy the shape of the array for each dimension. Also, the memory layout (linearization) of the array can be different. That means the array could be stored in row-major or column-major order - if we pass in the array strides correctly the kernel does the right thing. If you're not familiar with the concept of strides check out [this stackoverflow post](https://stackoverflow.com/questions/53097952/how-to-understand-numpy-strides-for-layman) or search in the numpy documentation for strides - C vs Fortran order.
 
 The goal of *pystencils* is to produce the fastest possible code. One technique to do this is to use all available information already on compile time and generate code that is highly adapted to the specific problem. In our case we already know the shape and strides of the arrays we want to apply the kernel to, so we can make use of this information. This idea leads to the **fixed-field-size workflow**. The main difference there is that we define the arrays first and therefore let *pystencils* know about the array shapes and strides, so that it can generate more specific code:
 
 1. create numpy arrays that hold your data
 2. define pystencils fields, this time telling pystencils already which arrays they correspond to, so that it knows about the size and strides
 
 in the other steps nothing has changed
 
 3. define the update rule
 4. compile update rule to kernel
 5. run the kernel
 
 %% Cell type:code id: tags:
 
 ``` python
 # 1. create arrays first
 src_arr, dst_arr = np.random.rand(30, 30), np.zeros([30, 30])
 
 # 2. define symbolic fields - note the additional parameter that link an array to each field
 src_field, dst_field = ps.fields("src, dst:[2D]", src=src_arr, dst=dst_arr)
 
 # 3. define update rule
 update_rule = [ps.Assignment(lhs=dst_field[0, 0],
                              rhs=(src_field[1, 0] + src_field[-1, 0] +
                                   src_field[0, 1] + src_field[0, -1]) / 4)]
 
 # 4. compile it
 kernel_function = ps.create_kernel(update_rule).compile()
 
 # 5. call kernel
 kernel_function(src=src_arr, dst=dst_arr)  # names of arguments have to match names passed to ps.fields()
 ```
 
 %% Cell type:markdown id: tags:
 
 Functionally, both variants are equivalent. We see the difference only when we look at the generated code
 
 %% Cell type:code id: tags:
 
 ``` python
 ps.show_code(kernel_function.ast)
 ```
 
 %% Output
 
 
 
 %% Cell type:markdown id: tags:
 
 Compare this to the code above! It looks much simpler. The reason is that all index computations are already simplified since the exact field sizes and strides are known. This kernel now only works on arrays of the previously specified size.
 
 Lets try what happens if we use a different array:
 
 %% Cell type:code id: tags:
 
 ``` python
 src_arr2, dst_arr2 = np.random.rand(40, 40), np.zeros([40, 40])
 try:
     kernel_function(src=src_arr2, dst=dst_arr2)
 except ValueError as e:
     print(e)
 ```
 
 %% Output
 
     Wrong shape of array dst. Expected (30, 30)
 
 %% Cell type:markdown id: tags:
 
 ### 1.2. GPU simulations
 
 Let's now jump to a seemingly unrelated topic: running kernels on the GPU.
-When creating the kernel, an additional parameter `target=ps.Target.GPU` has to be passed. Also, the compiled kernel cannot be called with numpy arrays directly, but has to be called with `pycuda.gpuarray`s instead. That means, we have to transfer our numpy array to GPU first. From this step we obtain a gpuarray, then we can run the kernel, hopefully multiple times so that the data transfer was worth the time. Finally we transfer the finished result back to CPU:
+When creating the kernel, an additional parameter `target=ps.Target.GPU` has to be passed. Also, the compiled kernel cannot be called with numpy arrays directly, but has to be called with `cupy.array`s instead. That means, we have to transfer our numpy array to GPU first. From this step we obtain a gpuarray, then we can run the kernel, hopefully multiple times so that the data transfer was worth the time. Finally we transfer the finished result back to CPU:
 
 %% Cell type:code id: tags:
 
 ``` python
-if pycuda:
+if cupy:
     config = ps.CreateKernelConfig(target=ps.Target.GPU)
     kernel_function_gpu = ps.create_kernel(update_rule, config=config).compile()
     # transfer to GPU
-    src_arr_gpu = pycuda.gpuarray.to_gpu(src_arr)
-    dst_arr_gpu = pycuda.gpuarray.to_gpu(dst_arr)
+    src_arr_gpu = cupy.asarray(src_arr)
+    dst_arr_gpu = cupy.asarray(dst_arr)
 
     # run kernel on GPU, this is done many times in real setups
     kernel_function_gpu(src=src_arr_gpu, dst=dst_arr_gpu)
 
     # transfer result back to CPU
-    dst_arr_gpu.get(dst_arr)
+    dst_arr = dst_arr_gpu.get()
 ```
 
 %% Cell type:markdown id: tags:
 
 ###  1.3. Summary: manual way
 
 - Don't confuse *pystencils* fields and *numpy* arrays
     - fields are symbolic
     - arrays are numeric
 - Use the fixed-field-size workflow whenever possible, since code might be faster. Create arrays first, then create fields from arrays
 - if we run GPU kernels, arrays have to transferred to the GPU first
 
 As demonstrated in the examples above we have to define 2 or 3 corresponding objects for each grid:
 
 - symbolic pystencils field
 - numpy array on CPU
-- for GPU run also a pycuda.gpuarray to mirror the data on the GPU
+- for GPU use also cupy to mirror the data on the GPU
 
 Managing these three objects manually is tedious and error-prone. We'll see in the next section how the data handling object takes care of this problem.
 
 %% Cell type:markdown id: tags:
 
 ## 2. Introducing the data handling - serial version
 
 ### 2.1. Example for CPU simulations
 
 The data handling internally keeps a mapping between symbolic fields and numpy arrays. When we create a field, automatically a corresponding array is allocated as well. Optionally we can also allocate memory on the GPU for the array as well. Lets dive right in and see how our example looks like, when implemented with a data handling.
 
 %% Cell type:code id: tags:
 
 ``` python
 dh = ps.create_data_handling(domain_size=(30, 30))
 
 # fields are now created using the data handling
 src_field = dh.add_array('src', values_per_cell=1)
 dst_field = dh.add_array('dst', values_per_cell=1)
 
 # kernel is created just like before
 update_rule = [ps.Assignment(lhs=dst_field[0, 0],
                              rhs=(src_field[1, 0] + src_field[-1, 0] + src_field[0, 1] + src_field[0, -1]) / 4)]
 kernel_function = ps.create_kernel(update_rule).compile()
 
 # have a look at the generated code - it uses
 # the fast version where array sizes are compiled-in
 #  ps.show_code(kernel_function.ast)
 ```
 
 %% Cell type:markdown id: tags:
 
 The data handling has methods to create fields - but where are the corresponding arrays?
 In the serial case you can access them as a member of the data handling, for example to initialize our 'src' array we can write
 
 %% Cell type:code id: tags:
 
 ``` python
 src_arr = dh.cpu_arrays['src']
 src_arr.fill(0.0)
 ```
 
 %% Cell type:markdown id: tags:
 
 This method is nice and easy, but you should not use it if you want your simulation to run on distributed-memory clusters. We'll see why in the last section. So it is good habit to not access the arrays directly but use the data handling to do so. We can, for example, initialize the array also with the following code:
 
 %% Cell type:code id: tags:
 
 ``` python
 dh.fill('src', 0.0)
 ```
 
 %% Cell type:markdown id: tags:
 
 To run the kernels with the same code as before, we would also need the arrays. We could do that accessing the `cpu_arrays`:
 
 %% Cell type:code id: tags:
 
 ``` python
 kernel_function(src=dh.cpu_arrays['src'],
                 dst=dh.cpu_arrays['dst'])
 ```
 
 %% Cell type:markdown id: tags:
 
 but to be prepared for MPI parallel simulations, again a method of the data handling should be used for this.
 Besides, this method is also simpler to use - since it automatically detects which arrays a kernel uses and passes them in.
 
 %% Cell type:code id: tags:
 
 ``` python
 dh.run_kernel(kernel_function)
 ```
 
 %% Cell type:markdown id: tags:
 
 To access the data read-only instead of using `cpu_arrays` the gather function should be used.
 This function gives you a read-only copy of the domain or part of the domain.
 We will discuss this function later in more detail when we look at MPI parallel simulations.
 For serial simulations keep in mind that modifying the resulting array does not change your original data!
 
 %% Cell type:code id: tags:
 
 ``` python
 read_only_copy = dh.gather_array('src', ps.make_slice[:, :], ghost_layers=False)
 ```
 
 %% Cell type:markdown id: tags:
 
 ### 2.2. Example for GPU simulations
 
 In this section we have a look at GPU simulations using the data handling. Only minimal changes are required.
 When creating the data handling we can pass a 'default_target'. This means for every added field an array on the CPU and the GPU is allocated. This is a useful default, for more fine-grained control the `add_array` method also takes additional parameters controlling where the array should be allocated.
 
 Additionally we also need to compile a GPU version of the kernel.
 
 %% Cell type:code id: tags:
 
 ``` python
 if gpu is False:
     dh = ps.create_data_handling(domain_size=(30, 30), default_target=ps.Target.CPU)
 else:
     dh = ps.create_data_handling(domain_size=(30, 30), default_target=ps.Target.GPU)
 
 # fields are now created using the data handling
 src_field = dh.add_array('src', values_per_cell=1)
 dst_field = dh.add_array('dst', values_per_cell=1)
 
 # kernel is created just like before
 update_rule = [ps.Assignment(lhs=dst_field[0, 0],
                              rhs=(src_field[1, 0] + src_field[-1, 0] + src_field[0, 1] + src_field[0, -1]) / 4)]
 config = ps.CreateKernelConfig(target=dh.default_target)
 kernel_function = ps.create_kernel(update_rule, config=config).compile()
 
 dh.fill('src', 0.0)
 ```
 
 %% Cell type:markdown id: tags:
 
 The data handling provides function to transfer data between CPU and GPU
 
 %% Cell type:code id: tags:
 
 ``` python
 if gpu:
     dh.to_gpu('src')
 
 dh.run_kernel(kernel_function)
 
 if gpu:
     dh.to_cpu('dst')
 ```
 
 %% Cell type:markdown id: tags:
 
 usually one wants to transfer all fields that have been allocated on CPU and GPU at the same time:
 
 %% Cell type:code id: tags:
 
 ``` python
 dh.all_to_gpu()
 dh.run_kernel(kernel_function)
 dh.all_to_cpu()
 ```
 
 %% Cell type:markdown id: tags:
 
 We can always include the `all_to_*` functions in our code, since they do nothing if there are no arrays allocated on the GPU. Thus there is only a single point in the code where we can switch between CPU and GPU version: the `default_target` of the data handling.
 
 %% Cell type:markdown id: tags:
 
 ### 2.2 Ghost Layers and periodicity
 
 The data handling can also provide periodic boundary conditions. Therefor the domain is extended by one layer of cells, the so-called ghost layer or halo layer.
 
 %% Cell type:code id: tags:
 
 ``` python
 print("Shape of domain         ", dh.shape)
 print("Direct access to arrays ", dh.cpu_arrays['src'].shape)
 print("Gather                  ", dh.gather_array('src', ghost_layers=True).shape)
 ```
 
 %% Output
 
     Shape of domain          (30, 30)
     Direct access to arrays  (32, 32)
     Gather                   (32, 32)
 
 %% Cell type:markdown id: tags:
 
  So the actual arrays are 2 cells larger than what you asked for. This additional layer is used to copy over the data from the other end of the array, such that for the stencil algorithm effectively the domain is periodic. This copying operation has to be started manually though:
 
 %% Cell type:code id: tags:
 
 ``` python
 dh = ps.create_data_handling((4, 4), periodicity=(True, False))
 dh.add_array("src")
 
 # get copy function
 copy_fct = dh.synchronization_function(['src'])
 
 dh.fill('src', 0.0, ghost_layers=True)
 dh.fill('src', 3.0, ghost_layers=False)
 before_sync = dh.gather_array('src', ghost_layers=True).copy()
 
 copy_fct() # copy over to get periodicity in x direction
 
 after_sync = dh.gather_array('src', ghost_layers=True).copy()
 plt.subplot(1,2,1)
 plt.scalar_field(before_sync);
 plt.title("Before")
 plt.subplot(1,2,2)
 plt.scalar_field(after_sync);
 plt.title("After");
 ```
 
 %% Output
 
-
+
 
 %% Cell type:markdown id: tags:
 
 ## 3. Going (MPI) parallel - the parallel data handling
 
 ### 3.1. Conceptual overview
 To run MPI parallel simulations the waLBerla framework is used. waLBerla has to be compiled against your local MPI library and thus is a bit hard to install. We suggest to use the version shipped with conda for testing. For production, when you want to run on a cluster the best option is to build waLBerla yourself against the MPI library that is installed on your cluster.
 
 Now lets have a look on how to write code that runs MPI parallel.
 Since the data is distributed, we don't have access to the full array any more but only to the part that is stored locally. The domain is split into so called blocks, where one process might get one (or sometimes multiple) blocks. To do anything with the local part of the data we iterate over the **local** blocks to get the contents as numpy arrays. The blocks returned in the loop differ from process to process.
 
 Copy the following snippet to a python file and run with multiple processes e.g.:
 ``mpirun -np 4 myscript.py`` you will see that there are as many blocks as processes.
 
 %% Cell type:code id: tags:
 
 ``` python
 dh = ps.create_data_handling(domain_size=(30, 30), parallel=True)
 field = dh.add_array('field')
 for block in dh.iterate():
     # offset is in global coordinates, where first inner cell has coordiante (0,0)
     # and ghost layers have negative coordinates
     print(block.offset, block['field'].shape)
 
     # use offset to create a local array 'my_data' for the part of the domain
     #np.copyto(block[field.name], my_data)
 ```
 
 %% Output
 
     (-1, -1) (32, 32)
+    [0][PROGRESS]------(0.000 sec) Initializing SetupBlockForest:
+    [0]                             - AABB: [ <0,0,0>, <30,30,1> ]
+    [0]                             - forest size (root blocks / blocks on the initial grid): 1 x 1 x 1
+    [0]                             - periodicity: false x false x false
+    [0][PROGRESS]------(0.000 sec) Initializing SetupBlockForest: Allocating root blocks ...
+    [0][PROGRESS]------(0.000 sec) Initializing SetupBlockForest: Setting up neighborhood information for each block (with respect to periodicity) ...
+    [0][PROGRESS]------(0.000 sec) Initializing SetupBlockForest: Assigning workload, memory requirements, and SUIDs to blocks ...
+    [0][PROGRESS]------(0.000 sec) Initializing SetupBlockForest: finished!
+    [0]                            The following block structure has been created:
+    [0]                            - AABB: [ <0,0,0>, <30,30,1> ]
+    [0]                            - initial decomposition: 1 x 1 x 1 (= forest size)
+    [0]                            - periodicity: false x false x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 1 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 1
+    [0]                            - blocks have not yet been distributed to processes
+    [0][PROGRESS]------(0.000 sec) Balancing SetupBlockForest: Creating a process distribution for 1 process(es) ...
+    [0][PROGRESS]------(0.000 sec) Balancing SetupBlockForest: process distribution to 1 process(es) finished!
+    [0]                            - number of worker processes:       1
+    [0]                            - number of empty buffer processes: 0
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            The resulting block structure looks like as follows:
+    [0]                            - AABB: [ <0,0,0>, <30,30,1> ]
+    [0]                            - initial decomposition: 1 x 1 x 1 (= forest size)
+    [0]                            - periodicity: false x false x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 1 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 1
+    [0]                            - number of processes: 1 (1 worker process(es) / 0 empty buffer process(es))
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            - process ID bytes: 0
+    [0]                            - blocks/memory/workload per process:
+    [0]                               + blocks:
+    [0]                                  - min       = 1
+    [0]                                  - max       = 1
+    [0]                                  - avg       = 1
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + memory:
+    [0]                                  - min       = 1
+    [0]                                  - max       = 1
+    [0]                                  - avg       = 1
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + workload:
+    [0]                                  - min       = 1
+    [0]                                  - max       = 1
+    [0]                                  - avg       = 1
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0][PROGRESS]------(0.000 sec) Adding block data ("cell bounding box")
+    [0][PROGRESS]------(0.000 sec) Adding block data ("field")
 
 %% Cell type:markdown id: tags:
 
 To get some more interesting results here in the notebook we put multiple blocks onto our single notebook process. This makes not much sense for real simulations, but for testing and demonstration purposes this is useful.
 
 %% Cell type:code id: tags:
 
 ``` python
 from waLBerla import createUniformBlockGrid
 from pystencils.datahandling import ParallelDataHandling
 
 blocks = createUniformBlockGrid(blocks=(2,1,2), cellsPerBlock=(20, 10, 20),
                                 oneBlockPerProcess=False, periodic=(1, 0, 0))
 dh = ParallelDataHandling(blocks)
 field = dh.add_array('field')
 for block in dh.iterate():
     print(block.offset, block['field'].shape)
 ```
 
 %% Output
 
-    (-1, -1, -1) (22, 12, 22)
+    (-1, -1, -1)[0][PROGRESS]------(0.025 sec) Initializing SetupBlockForest:
+    [0]                             - AABB: [ <0,0,0>, <40,10,40> ]
+    [0]                             - forest size (root blocks / blocks on the initial grid): 2 x 1 x 2
+    [0]                             - periodicity: true x false x false
+     (22, 12, 22)
     (19, -1, -1) (22, 12, 22)
     (-1, -1, 19) (22, 12, 22)
     (19, -1, 19) (22, 12, 22)
+    [0][PROGRESS]------(0.025 sec) Initializing SetupBlockForest: Allocating root blocks ...
+    [0][PROGRESS]------(0.025 sec) Initializing SetupBlockForest: Setting up neighborhood information for each block (with respect to periodicity) ...
+    [0][PROGRESS]------(0.025 sec) Initializing SetupBlockForest: Assigning workload, memory requirements, and SUIDs to blocks ...
+    [0][PROGRESS]------(0.025 sec) Initializing SetupBlockForest: finished!
+    [0]                            The following block structure has been created:
+    [0]                            - AABB: [ <0,0,0>, <40,10,40> ]
+    [0]                            - initial decomposition: 2 x 1 x 2 (= forest size)
+    [0]                            - periodicity: true x false x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 3 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 4
+    [0]                            - blocks have not yet been distributed to processes
+    [0][PROGRESS]------(0.025 sec) Balancing SetupBlockForest: Creating a process distribution for 1 process(es) ...
+    [0][PROGRESS]------(0.025 sec) Balancing SetupBlockForest: process distribution to 1 process(es) finished!
+    [0]                            - number of worker processes:       1
+    [0]                            - number of empty buffer processes: 0
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            The resulting block structure looks like as follows:
+    [0]                            - AABB: [ <0,0,0>, <40,10,40> ]
+    [0]                            - initial decomposition: 2 x 1 x 2 (= forest size)
+    [0]                            - periodicity: true x false x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 3 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 4
+    [0]                            - number of processes: 1 (1 worker process(es) / 0 empty buffer process(es))
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            - process ID bytes: 0
+    [0]                            - blocks/memory/workload per process:
+    [0]                               + blocks:
+    [0]                                  - min       = 4
+    [0]                                  - max       = 4
+    [0]                                  - avg       = 4
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + memory:
+    [0]                                  - min       = 4
+    [0]                                  - max       = 4
+    [0]                                  - avg       = 4
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + workload:
+    [0]                                  - min       = 4
+    [0]                                  - max       = 4
+    [0]                                  - avg       = 4
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0][PROGRESS]------(0.025 sec) Adding block data ("cell bounding box")
+    [0][PROGRESS]------(0.026 sec) Adding block data ("field")
 
 %% Cell type:markdown id: tags:
 
 Now we see that we have four blocks with (20, 10, 20) block each, and the global domain is (40, 10, 40) big.
 All subblock also have a ghost layer around them, which is used to synchronize with their neighboring blocks (over the network). For ghost layer synchronization the same `synchronization_function` is used that we used above for periodic boundaries, because copying between blocks and copying the ghost layer for periodicity uses the same mechanism.
 
 %% Cell type:code id: tags:
 
 ``` python
 dh.gather_array('field').shape
 ```
 
 %% Output
 
-
+
     $\displaystyle \left( 40, \  10, \  40\right)$
     (40, 10, 40)
 
 %% Cell type:markdown id: tags:
 
 ### 3.2. Parallel example
 
 To illustrate this in more detail, lets run a simple kernel on a parallel domain. waLBerla can handle 3D domains only, so we choose a z-size of 1.
 
 %% Cell type:code id: tags:
 
 ``` python
 blocks = createUniformBlockGrid(blocks=(2,4,1), cellsPerBlock=(20, 10, 1),
                                 oneBlockPerProcess=False, periodic=(1, 1, 0))
 dh = ParallelDataHandling(blocks, dim=2)
 src_field = dh.add_array('src')
 dst_field = dh.add_array('dst')
 
 update_rule = [ps.Assignment(lhs=dst_field[0, 0 ],
                              rhs=(src_field[1, 0] + src_field[-1, 0] +
                                   src_field[0, 1] + src_field[0, -1]) / 4)]
 
 # 3. compile update rule to function
 kernel_function = ps.create_kernel(update_rule).compile()
 ```
 
+%% Output
+
+    [0][PROGRESS]------(0.119 sec) Initializing SetupBlockForest:
+    [0]                             - AABB: [ <0,0,0>, <40,40,1> ]
+    [0]                             - forest size (root blocks / blocks on the initial grid): 2 x 4 x 1
+    [0]                             - periodicity: true x true x false
+    [0][PROGRESS]------(0.119 sec) Initializing SetupBlockForest: Allocating root blocks ...
+    [0][PROGRESS]------(0.119 sec) Initializing SetupBlockForest: Setting up neighborhood information for each block (with respect to periodicity) ...
+    [0][PROGRESS]------(0.119 sec) Initializing SetupBlockForest: Assigning workload, memory requirements, and SUIDs to blocks ...
+    [0][PROGRESS]------(0.119 sec) Initializing SetupBlockForest: finished!
+    [0]                            The following block structure has been created:
+    [0]                            - AABB: [ <0,0,0>, <40,40,1> ]
+    [0]                            - initial decomposition: 2 x 4 x 1 (= forest size)
+    [0]                            - periodicity: true x true x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 4 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 8
+    [0]                            - blocks have not yet been distributed to processes
+    [0][PROGRESS]------(0.119 sec) Balancing SetupBlockForest: Creating a process distribution for 1 process(es) ...
+    [0][PROGRESS]------(0.119 sec) Balancing SetupBlockForest: process distribution to 1 process(es) finished!
+    [0]                            - number of worker processes:       1
+    [0]                            - number of empty buffer processes: 0
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            The resulting block structure looks like as follows:
+    [0]                            - AABB: [ <0,0,0>, <40,40,1> ]
+    [0]                            - initial decomposition: 2 x 4 x 1 (= forest size)
+    [0]                            - periodicity: true x true x false
+    [0]                            - number of blocks discarded from the initial grid: 0 (= 0 %)
+    [0]                            - number of levels: 1
+    [0]                            - tree ID digits: 4 (-> block ID bytes = 1)
+    [0]                            - total number of blocks: 8
+    [0]                            - number of processes: 1 (1 worker process(es) / 0 empty buffer process(es))
+    [0]                            - buffer processes are inserted into the process network: no
+    [0]                            - process ID bytes: 0
+    [0]                            - blocks/memory/workload per process:
+    [0]                               + blocks:
+    [0]                                  - min       = 8
+    [0]                                  - max       = 8
+    [0]                                  - avg       = 8
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + memory:
+    [0]                                  - min       = 8
+    [0]                                  - max       = 8
+    [0]                                  - avg       = 8
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0]                               + workload:
+    [0]                                  - min       = 8
+    [0]                                  - max       = 8
+    [0]                                  - avg       = 8
+    [0]                                  - stdDev    = 0
+    [0]                                  - relStdDev = 0
+    [0][PROGRESS]------(0.119 sec) Adding block data ("cell bounding box")
+    [0][PROGRESS]------(0.119 sec) Adding block data ("src")
+    [0][PROGRESS]------(0.120 sec) Adding block data ("dst")
+
 %% Cell type:markdown id: tags:
 
 Now lets initialize the arrays. To do this we can get arrays (meshgrids) from the block with the coordinates of the local cells. We use this to initialize a circular shape.
 
 %% Cell type:code id: tags:
 
 ``` python
 dh.fill('src', 0.0)
 for block in dh.iterate(ghost_layers=False, inner_ghost_layers=False):
     x, y = block.midpoint_arrays
     inside_sphere = (x -20)**2 + (y-25)**2  < 8 ** 2
     block['src'][inside_sphere] = 1.0
 plt.scalar_field( dh.gather_array('src') );
 ```
 
 %% Output
 
-
+
 
 %% Cell type:markdown id: tags:
 
 Now we can run our compute kernel on this data as usual. We just have to make sure that the field is synchronized after every step, i.e. that the ghost layers are correctly updated.
 
 %% Cell type:code id: tags:
 
 ``` python
 sync = dh.synchronization_function(['src'])
 for i in range(40):
     sync()
     dh.run_kernel(kernel_function)
     dh.swap('src', 'dst')
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 plt.scalar_field( dh.gather_array('src') );
 ```
 
 %% Output
 
-
+

doc/notebooks/demo_assignment_collection.ipynb

 %% Cell type:code id: tags:
 
 ``` python
 from pystencils.session import *
 ```
 
 %% Cell type:markdown id: tags:
 
 # Demo: Assignment collections and simplification
 
 
 ## Assignment collections
 
 The assignment collection class helps to formulate and simplify assignments for numerical kernels.
 
 An ``AssignmentCollection`` is an ordered collection of assignments, together with an optional ordered collection of subexpressions, that are required to evaluate the main assignments. There are various simplification rules available that operate on ``AssignmentCollection``s.
 
 %% Cell type:markdown id: tags:
 
 We start by defining some stencil update rule. Here we also use the *pystencils* ``Field``, note however that the assignment collection module works purely on the *sympy* level.
 
 %% Cell type:code id: tags:
 
 ``` python
 a,b,c = sp.symbols("a b c")
 f = ps.fields("f(2) : [2D]")
+g = ps.fields("g(2) : [2D]")
 
-a1 = ps.Assignment(f[0,0](1), (a**2 +b) * f[0,1] + \
+a1 = ps.Assignment(g[0,0](1), (a**2 +b) * f[0,1] + \
                   (a**2 - c) * f[1,0] + \
                   (a**2 - 2*c) * f[-1,0] + \
                   (a**2) * f[0, -1])
 
-a2 = ps.Assignment(f[0,0](0), (c**2 +b) * f[0,1] + \
+a2 = ps.Assignment(g[0,0](0), (c**2 +b) * f[0,1] + \
                   (c**2 - c) * f[1,0] + \
                   (c**2 - 2*c) * f[-1,0] + \
                   (c**2 - a**2) * f[0, -1])
 
 
 ac = ps.AssignmentCollection([a1, a2], subexpressions=[])
 ac
 ```
 
 %% Output
 
-    Equation Collection for f_C^1,f_C^0
+    AssignmentCollection: g_C^0, g_C^1 <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 *sympy* operations can be applied on an assignment collection: In this example we first expand the collection, then look for common subexpressions.
 
 %% Cell type:code id: tags:
 
 ``` python
 expand_all = ps.simp.apply_to_all_assignments(sp.expand)
 expandedEc = expand_all(ac)
 ```
 
 %% Cell type:code id: tags:
 
 ``` python
 ac_cse = ps.simp.sympy_cse(expandedEc)
 ac_cse
 ```
 
 %% Output
 
-    Equation Collection for f_C^1,f_C^0
+    AssignmentCollection: g_C^0, g_C^1 <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 Symbols occuring in assignment collections are classified into 3 categories:
 - ``free_symbols``: symbols that occur in right-hand-sides but never on left-hand-sides
 - ``bound_symbols``: symbols that occur on left-hand-sides
 - ``defined_symbols``: symbols that occur on left-hand-sides of a main assignment
 
 %% Cell type:code id: tags:
 
 ``` python
 ac_cse.free_symbols
 ```
 
 %% Output
 
-
-    $$\left\{{{f}_{E}^{0}}, {{f}_{N}^{0}}, {{f}_{S}^{0}}, {{f}_{W}^{0}}, a, b, c\right\}$$
-    set([f_E__0, f_N__0, f_S__0, f_W__0, a, b, c])
+
+    $\displaystyle \left\{{f}_{(1,0)}^{0}, {f}_{(0,1)}^{0}, {f}_{(0,-1)}^{0}, {f}_{(-1,0)}^{0}, a, b, c\right\}$
+    {f_E__0, f_N__0, f_S__0, f_W__0, a, b, c}
 
 %% Cell type:code id: tags:
 
 ``` python
 ac_cse.bound_symbols
 ```
 
 %% Output
 
-
-    $$\left\{{{f}_{C}^{0}}, {{f}_{C}^{1}}, \xi_{0}, \xi_{1}, \xi_{2}, \xi_{3}\right\}$$
-    set([f_C__0, f_C__1, ξ₀, ξ₁, ξ₂, ξ₃])
+
+    $\displaystyle \left\{{g}_{(0,0)}^{0}, {g}_{(0,0)}^{1}, \xi_{0}, \xi_{1}, \xi_{2}, \xi_{3}\right\}$
+    {g_C__0, g_C__1, ξ₀, ξ₁, ξ₂, ξ₃}
 
 %% Cell type:code id: tags:
 
 ``` python
 ac_cse.defined_symbols
 ```
 
 %% Output
 
-
-    $$\left\{{{f}_{C}^{0}}, {{f}_{C}^{1}}\right\}$$
-    set([f_C__0, f_C__1])
+
+    $\displaystyle \left\{{g}_{(0,0)}^{0}, {g}_{(0,0)}^{1}\right\}$
+    {g_C__0, g_C__1}
 
 %% Cell type:markdown id: tags:
 
 Assignment collections can be splitted up, and merged together. For splitting, a list of symbols that occur on the left-hand-side in the main assignments has to be passed. The returned assignment collection only contains these main assignments together with all necessary subexpressions.
 
 %% Cell type:code id: tags:
 
 ``` python
-ac_f0 = ac_cse.new_filtered([f(0)])
-ac_f1 = ac_cse.new_filtered([f(1)])
+ac_f0 = ac_cse.new_filtered([g(0)])
+ac_f1 = ac_cse.new_filtered([g(1)])
 ac_f1
 ```
 
 %% Output
 
-    Equation Collection for f_C^1
+    AssignmentCollection: g_C^1, <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 Note here that $\xi_4$ is no longer part of the subexpressions, since it is not used in the main assignment of $f_C^1$.
 
 If we merge both collections together, we end up with the original collection.
 
 %% Cell type:code id: tags:
 
 ``` python
 ac_f0.new_merged(ac_f1)
 ```
 
 %% Output
 
-    Equation Collection for f_C^0,f_C^1
+    AssignmentCollection: g_C^0, g_C^1 <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 There is also a method that inserts all subexpressions into the main assignments. This is the inverse operation of common subexpression elimination.
 
 %% Cell type:code id: tags:
 
 ``` python
 assert sp.simplify(ac_f0.new_without_subexpressions().main_assignments[0].rhs - a2.rhs) == 0
 ac_f0.new_without_subexpressions()
 ```
 
 %% Output
 
-    Equation Collection for f_C^0
+    AssignmentCollection: g_C^0, <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 To evaluate an assignment collection, use the ``lambdify`` method. It is very similar to *sympy*s ``lambdify`` function.
 
 %% Cell type:code id: tags:
 
 ``` python
 evalFct = ac_cse.lambdify([f[0,1], f[1,0]],  # new parameters of returned function
                           fixed_symbols={a:1, b:2, c:3, f[0,-1]: 4, f[-1,0]: 5}) # fix values of other symbols
 evalFct(2,1)
 ```
 
 %% Output
 
-
-    $$\left \{ {{f}_{C}^{0}} : 75, \quad {{f}_{C}^{1}} : -17\right \}$$
-    {f_C__0: 75, f_C__1: -17}
+
+    $\displaystyle \left\{ {g}_{(0,0)}^{0} : 75, \  {g}_{(0,0)}^{1} : -17\right\}$
+    {g_C__0: 75, g_C__1: -17}
 
 %% Cell type:markdown id: tags:
 
 lambdify is rather slow for evaluation. The intended way to evaluate an assignment collection is *pystencils* i.e. create a fast kernel, that applies the update at every site of a structured grid. The collection can be directly passed to the `create_kernel` function.
 
 %% Cell type:code id: tags:
 
 ``` python
 func = ps.create_kernel(ac_cse).compile()
 ```
 
 %% Cell type:markdown id: tags:
 
 ## Simplification Strategies
 
 In above examples, we already applied simplification rules to assignment collections. Simplification rules are functions that take, as a single argument, an assignment collection and return an modified/simplified copy of it. The ``SimplificationStrategy`` class holds a list of simplification rules and can apply all of them in the specified order. Additionally it provides useful printing and reporting functions.
 
 We start by creating a simplification strategy, consisting of the expand and CSE simplifications we have already applied above:
 
 %% Cell type:code id: tags:
 
 ``` python
 strategy = ps.simp.SimplificationStrategy()
 strategy.add(ps.simp.apply_to_all_assignments(sp.expand))
 strategy.add(ps.simp.sympy_cse)
 ```
 
 %% Cell type:markdown id: tags:
 
 This strategy can be applied to any assignment collection:
 
 %% Cell type:code id: tags:
 
 ``` python
 strategy(ac)
 ```
 
 %% Output
 
-    Equation Collection for f_C^1,f_C^0
+    AssignmentCollection: g_C^0, g_C^1 <- f(f_N^0, b, f_S^0, f_E^0, a, f_W^0, c)
 
 %% Cell type:markdown id: tags:
 
 The strategy can also print the simplification results at each stage.
 The report contains information about the number of operations after each simplification as well as the runtime of each simplification routine.
 
 %% Cell type:code id: tags:
 
 ``` python
 strategy.create_simplification_report(ac)
 ```
 
 %% Output
 
-    <pystencils.simp.simplificationstrategy.SimplificationStrategy.create_simplification_report.<locals>.Report at 0x7f9be404fda0>
+    <pystencils.simp.simplificationstrategy.SimplificationStrategy.create_simplification_report.<locals>.Report at 0x147de3e90>
 
 %% Cell type:markdown id: tags:
 
 The strategy can also print the full collection after each simplification...
 
 %% Cell type:code id: tags:
 
 ``` python
 strategy.show_intermediate_results(ac)
 ```
 
 %% Output
 
-    <pystencils.simp.simplificationstrategy.SimplificationStrategy.show_intermediate_results.<locals>.IntermediateResults at 0x7f9bad688dd8>
+    <pystencils.simp.simplificationstrategy.SimplificationStrategy.show_intermediate_results.<locals>.IntermediateResults at 0x147e09c90>
 
 %% Cell type:markdown id: tags:
 
 ... or only specific assignments for better readability
 
 %% Cell type:code id: tags:
 
 ``` python
-strategy.show_intermediate_results(ac, symbols=[f(1)])
+strategy.show_intermediate_results(ac, symbols=[g(1)])
 ```
 
 %% Output
 
-    <pystencils.simp.simplificationstrategy.SimplificationStrategy.show_intermediate_results.<locals>.IntermediateResults at 0x7f9bad688b00>
+    <pystencils.simp.simplificationstrategy.SimplificationStrategy.show_intermediate_results.<locals>.IntermediateResults at 0x1265a1b90>
+
+%% Cell type:code id: tags:
+
+``` python
+```

doc/sphinx/kernel_compile_and_call.rst

@@ -27,11 +27,11 @@ Code printing
 GPU Indexing
 -------------
 
-.. autoclass:: pystencils.gpucuda.AbstractIndexing
+.. autoclass:: pystencils.gpu.AbstractIndexing
    :members:
 
-.. autoclass:: pystencils.gpucuda.BlockIndexing
+.. autoclass:: pystencils.gpu.BlockIndexing
    :members:
 
-.. autoclass:: pystencils.gpucuda.LineIndexing
+.. autoclass:: pystencils.gpu.LineIndexing
    :members:

pyproject.toml

+[project]
+name = "pystencils"
+description = "Speeding up stencil computations on CPUs and GPUs"
+dynamic = ["version"]
+readme = "README.md"
+authors = [
+    { name = "Martin Bauer" },
+    { name = "Jan Hönig " },
+    { name = "Markus Holzer" },
+    { name = "Frederik Hennig" },
+    { email = "cs10-codegen@fau.de" },
+]
+license = { file = "COPYING.txt" }
+requires-python = ">=3.10"
+dependencies = ["sympy>=1.9,<=1.12.1", "numpy>=1.8.0", "appdirs", "joblib", "pyyaml", "fasteners"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Framework :: Jupyter",
+    "Topic :: Software Development :: Code Generators",
+    "Topic :: Scientific/Engineering :: Physics",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: GNU Affero General Public License v3 or later (AGPLv3+)",
+]
+
+[project.urls]
+"Bug Tracker" = "https://i10git.cs.fau.de/pycodegen/pystencils/-/issues"
+"Documentation" = "https://pycodegen.pages.i10git.cs.fau.de/pystencils/"
+"Source Code" = "https://i10git.cs.fau.de/pycodegen/pystencils"
+
+[project.optional-dependencies]
+gpu = ['cupy']
+alltrafos = ['islpy', 'py-cpuinfo']
+bench_db = ['blitzdb', 'pymongo', 'pandas']
+interactive = [
+    'matplotlib',
+    'ipy_table',
+    'imageio',
+    'jupyter',
+    'pyevtk',
+    'rich',
+    'graphviz',
+]
+use_cython = [
+    'Cython'
+]
+doc = [
+    'sphinx',
+    'sphinx_rtd_theme',
+    'nbsphinx',
+    'sphinxcontrib-bibtex',
+    'sphinx_autodoc_typehints',
+    'pandoc',
+]
+tests = [
+    'pytest',
+    'pytest-cov',
+    'pytest-html',
+    'ansi2html',
+    'pytest-xdist',
+    'flake8',
+    'nbformat',
+    'nbconvert',
+    'ipython',
+    'matplotlib',
+    'py-cpuinfo',
+    'randomgen>=1.18',
+]
+
+[build-system]
+requires = [
+    "setuptools>=61",
+    "versioneer[toml]>=0.29",
+    # 'Cython'
+]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.package-data]
+pystencils = [
+    "include/*.h",
+    "boundaries/createindexlistcython.pyx"
+]
+
+[tool.setuptools.packages.find]
+where = ["src"]
+include = ["pystencils", "pystencils.*"]
+namespaces = false
+
+[tool.versioneer]
+# See the docstring in versioneer.py for instructions. Note that you must
+# re-run 'versioneer.py setup' after changing this section, and commit the
+# resulting files.
+VCS = "git"
+style = "pep440"
+versionfile_source = "src/pystencils/_version.py"
+versionfile_build = "pystencils/_version.py"
+tag_prefix = "release/"
+parentdir_prefix = "pystencils-"

pystencils/datahandling/pycuda.py

-try:
-    import pycuda.gpuarray as gpuarray
-except ImportError:
-    gpuarray = None
-import numpy as np
-
-import pystencils
-
-
-class PyCudaArrayHandler:
-
-    def __init__(self):
-        import pycuda.autoinit  # NOQA
-
-    def zeros(self, shape, dtype=np.float64, order='C'):
-        cpu_array = np.zeros(shape=shape, dtype=dtype, order=order)
-        return self.to_gpu(cpu_array)
-
-    def ones(self, shape, dtype=np.float64, order='C'):
-        cpu_array = np.ones(shape=shape, dtype=dtype, order=order)
-        return self.to_gpu(cpu_array)
-
-    def empty(self, shape, dtype=np.float64, layout=None):
-        if layout:
-            cpu_array = pystencils.field.create_numpy_array_with_layout(shape=shape, dtype=dtype, layout=layout)
-            return self.to_gpu(cpu_array)
-        else:
-            return gpuarray.empty(shape, dtype)
-
-    @staticmethod
-    def to_gpu(array):
-        return gpuarray.to_gpu(array)
-
-    @staticmethod
-    def upload(array, numpy_array):
-        array.set(numpy_array)
-
-    @staticmethod
-    def download(array, numpy_array):
-        array.get(numpy_array)
-
-    def randn(self, shape, dtype=np.float64):
-        cpu_array = np.random.randn(*shape).astype(dtype)
-        return self.to_gpu(cpu_array)
-
-    from_numpy = to_gpu
-
-
-class PyCudaNotAvailableHandler:
-    def __getattribute__(self, name):
-        raise NotImplementedError("Unable to initiaize PyCuda! "
-                                  "Try to run `import pycuda.autoinit` to check whether PyCuda is working correctly!")

pystencils/include/PyStencilsField.h

-#pragma once
-
-extern "C++" {
-#ifdef __CUDA_ARCH__
-template <typename DTYPE_T, std::size_t DIMENSION> struct PyStencilsField {
-  DTYPE_T *data;
-  DTYPE_T shape[DIMENSION];
-  DTYPE_T stride[DIMENSION];
-};
-#else
-#include <array>
-
-template <typename DTYPE_T, std::size_t DIMENSION> struct PyStencilsField {
-  DTYPE_T *data;
-  std::array<DTYPE_T, DIMENSION> shape;
-  std::array<DTYPE_T, DIMENSION> stride;
-};
-#endif
-}

pystencils/include/cuda_complex.hpp

-// An implementation of C++ std::complex for use on CUDA devices.
-// Written by John C. Travers <jtravs@gmail.com> (2012)
-//
-// Missing:
-//  - long double support (not supported on CUDA)
-//  - some integral pow functions (due to lack of C++11 support on CUDA)
-//
-// Heavily derived from the LLVM libcpp project (svn revision 147853).
-// Based on libcxx/include/complex.
-// The git history contains the complete change history from the original.
-// The modifications are licensed as per the original LLVM license below.
-//
-// -*- C++ -*-
-//===--------------------------- complex ----------------------------------===//
-//
-//                     The LLVM Compiler Infrastructure
-//
-// This file is dual licensed under the MIT and the University of Illinois Open
-// Source Licenses. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-
-extern "C++" {
-#ifndef CUDA_COMPLEX_HPP
-#define CUDA_COMPLEX_HPP
-
-#ifdef __CUDACC__
-#define CUDA_CALLABLE_MEMBER __host__ __device__
-#else
-#define CUDA_CALLABLE_MEMBER
-#endif
-
-/*
-    complex synopsis
-
-template<class T>
-class complex
-{
-public:
-    typedef T value_type;
-
-    complex(const T& re = T(), const T& im = T());
-    complex(const complex&);
-    template<class X> complex(const complex<X>&);
-
-    T real() const;
-    T imag() const;
-
-    void real(T);
-    void imag(T);
-
-    complex<T>& operator= (const T&);
-    complex<T>& operator+=(const T&);
-    complex<T>& operator-=(const T&);
-    complex<T>& operator*=(const T&);
-    complex<T>& operator/=(const T&);
-
-    complex& operator=(const complex&);
-    template<class X> complex<T>& operator= (const complex<X>&);
-    template<class X> complex<T>& operator+=(const complex<X>&);
-    template<class X> complex<T>& operator-=(const complex<X>&);
-    template<class X> complex<T>& operator*=(const complex<X>&);
-    template<class X> complex<T>& operator/=(const complex<X>&);
-};
-
-template<>
-class complex<float>
-{
-public:
-    typedef float value_type;
-
-    constexpr complex(float re = 0.0f, float im = 0.0f);
-    explicit constexpr complex(const complex<double>&);
-
-    constexpr float real() const;
-    void real(float);
-    constexpr float imag() const;
-    void imag(float);
-
-    complex<float>& operator= (float);
-    complex<float>& operator+=(float);
-    complex<float>& operator-=(float);
-    complex<float>& operator*=(float);
-    complex<float>& operator/=(float);
-
-    complex<float>& operator=(const complex<float>&);
-    template<class X> complex<float>& operator= (const complex<X>&);
-    template<class X> complex<float>& operator+=(const complex<X>&);
-    template<class X> complex<float>& operator-=(const complex<X>&);
-    template<class X> complex<float>& operator*=(const complex<X>&);
-    template<class X> complex<float>& operator/=(const complex<X>&);
-};
-
-template<>
-class complex<double>
-{
-public:
-    typedef double value_type;
-
-    constexpr complex(double re = 0.0, double im = 0.0);
-    constexpr complex(const complex<float>&);
-
-    constexpr double real() const;
-    void real(double);
-    constexpr double imag() const;
-    void imag(double);
-
-    complex<double>& operator= (double);
-    complex<double>& operator+=(double);
-    complex<double>& operator-=(double);
-    complex<double>& operator*=(double);
-    complex<double>& operator/=(double);
-    complex<double>& operator=(const complex<double>&);
-
-    template<class X> complex<double>& operator= (const complex<X>&);
-    template<class X> complex<double>& operator+=(const complex<X>&);
-    template<class X> complex<double>& operator-=(const complex<X>&);
-    template<class X> complex<double>& operator*=(const complex<X>&);
-    template<class X> complex<double>& operator/=(const complex<X>&);
-};
-
-// 26.3.6 operators:
-template<class T> complex<T> operator+(const complex<T>&, const complex<T>&);
-template<class T> complex<T> operator+(const complex<T>&, const T&);
-template<class T> complex<T> operator+(const T&, const complex<T>&);
-template<class T> complex<T> operator-(const complex<T>&, const complex<T>&);
-template<class T> complex<T> operator-(const complex<T>&, const T&);
-template<class T> complex<T> operator-(const T&, const complex<T>&);
-template<class T> complex<T> operator*(const complex<T>&, const complex<T>&);
-template<class T> complex<T> operator*(const complex<T>&, const T&);
-template<class T> complex<T> operator*(const T&, const complex<T>&);
-template<class T> complex<T> operator/(const complex<T>&, const complex<T>&);
-template<class T> complex<T> operator/(const complex<T>&, const T&);
-template<class T> complex<T> operator/(const T&, const complex<T>&);
-template<class T> complex<T> operator+(const complex<T>&);
-template<class T> complex<T> operator-(const complex<T>&);
-template<class T> bool operator==(const complex<T>&, const complex<T>&);
-template<class T> bool operator==(const complex<T>&, const T&);
-template<class T> bool operator==(const T&, const complex<T>&);
-template<class T> bool operator!=(const complex<T>&, const complex<T>&);
-template<class T> bool operator!=(const complex<T>&, const T&);
-template<class T> bool operator!=(const T&, const complex<T>&);
-
-template<class T, class charT, class traits>
-  basic_istream<charT, traits>&
-  operator>>(basic_istream<charT, traits>&, complex<T>&);
-template<class T, class charT, class traits>
-  basic_ostream<charT, traits>&
-  operator<<(basic_ostream<charT, traits>&, const complex<T>&);
-
-// 26.3.7 values:
-
-template<class T>              T real(const complex<T>&);
-                          double real(double);
-template<Integral T>      double real(T);
-                          float  real(float);
-
-template<class T>              T imag(const complex<T>&);
-                          double imag(double);
-template<Integral T>      double imag(T);
-                          float  imag(float);
-
-template<class T> T abs(const complex<T>&);
-
-template<class T>              T arg(const complex<T>&);
-                          double arg(double);
-template<Integral T>      double arg(T);
-                          float  arg(float);
-
-template<class T>              T norm(const complex<T>&);
-                          double norm(double);
-template<Integral T>      double norm(T);
-                          float  norm(float);
-
-template<class T>      complex<T>           conj(const complex<T>&);
-                       complex<double>      conj(double);
-template<Integral T>   complex<double>      conj(T);
-                       complex<float>       conj(float);
-
-template<class T>    complex<T>           proj(const complex<T>&);
-                     complex<double>      proj(double);
-template<Integral T> complex<double>      proj(T);
-                     complex<float>       proj(float);
-
-template<class T> complex<T> polar(const T&, const T& = 0);
-
-// 26.3.8 transcendentals:
-template<class T> complex<T> acos(const complex<T>&);
-template<class T> complex<T> asin(const complex<T>&);
-template<class T> complex<T> atan(const complex<T>&);
-template<class T> complex<T> acosh(const complex<T>&);
-template<class T> complex<T> asinh(const complex<T>&);
-template<class T> complex<T> atanh(const complex<T>&);
-template<class T> complex<T> cos (const complex<T>&);
-template<class T> complex<T> cosh (const complex<T>&);
-template<class T> complex<T> exp (const complex<T>&);
-template<class T> complex<T> log (const complex<T>&);
-template<class T> complex<T> log10(const complex<T>&);
-
-template<class T> complex<T> pow(const complex<T>&, const T&);
-template<class T> complex<T> pow(const complex<T>&, const complex<T>&);
-template<class T> complex<T> pow(const T&, const complex<T>&);
-
-template<class T> complex<T> sin (const complex<T>&);
-template<class T> complex<T> sinh (const complex<T>&);
-template<class T> complex<T> sqrt (const complex<T>&);
-template<class T> complex<T> tan (const complex<T>&);
-template<class T> complex<T> tanh (const complex<T>&);
-
-template<class T, class charT, class traits>
-  basic_istream<charT, traits>&
-  operator>>(basic_istream<charT, traits>& is, complex<T>& x);
-
-template<class T, class charT, class traits>
-  basic_ostream<charT, traits>&
-  operator<<(basic_ostream<charT, traits>& o, const complex<T>& x);
-
-*/
-
-#include <math.h>
-#include <sstream>
-
-template <class _Tp> class complex;
-
-template <class _Tp>
-complex<_Tp> operator*(const complex<_Tp> &__z, const complex<_Tp> &__w);
-template <class _Tp>
-complex<_Tp> operator/(const complex<_Tp> &__x, const complex<_Tp> &__y);
-
-template <class _Tp> class complex {
-public:
-  typedef _Tp value_type;
-
-private:
-  value_type __re_;
-  value_type __im_;
-
-public:
-  CUDA_CALLABLE_MEMBER
-  complex(const value_type &__re = value_type(),
-          const value_type &__im = value_type())
-      : __re_(__re), __im_(__im) {}
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex(const complex<_Xp> &__c)
-      : __re_(__c.real()), __im_(__c.imag()) {}
-
-  CUDA_CALLABLE_MEMBER value_type real() const { return __re_; }
-  CUDA_CALLABLE_MEMBER value_type imag() const { return __im_; }
-
-  CUDA_CALLABLE_MEMBER void real(value_type __re) { __re_ = __re; }
-  CUDA_CALLABLE_MEMBER void imag(value_type __im) { __im_ = __im; }
-
-  CUDA_CALLABLE_MEMBER complex &operator=(const value_type &__re) {
-    __re_ = __re;
-    __im_ = value_type();
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator+=(const value_type &__re) {
-    __re_ += __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator-=(const value_type &__re) {
-    __re_ -= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator*=(const value_type &__re) {
-    __re_ *= __re;
-    __im_ *= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator/=(const value_type &__re) {
-    __re_ /= __re;
-    __im_ /= __re;
-    return *this;
-  }
-
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator=(const complex<_Xp> &__c) {
-    __re_ = __c.real();
-    __im_ = __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator+=(const complex<_Xp> &__c) {
-    __re_ += __c.real();
-    __im_ += __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator-=(const complex<_Xp> &__c) {
-    __re_ -= __c.real();
-    __im_ -= __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator*=(const complex<_Xp> &__c) {
-    *this = *this * __c;
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator/=(const complex<_Xp> &__c) {
-    *this = *this / __c;
-    return *this;
-  }
-};
-
-template <> class complex<double>;
-
-template <> class complex<float> {
-  float __re_;
-  float __im_;
-
-public:
-  typedef float value_type;
-
-  /*constexpr*/ CUDA_CALLABLE_MEMBER complex(float __re = 0.0f,
-                                             float __im = 0.0f)
-      : __re_(__re), __im_(__im) {}
-  explicit /*constexpr*/ complex(const complex<double> &__c);
-
-  /*constexpr*/ CUDA_CALLABLE_MEMBER float real() const { return __re_; }
-  /*constexpr*/ CUDA_CALLABLE_MEMBER float imag() const { return __im_; }
-
-  CUDA_CALLABLE_MEMBER void real(value_type __re) { __re_ = __re; }
-  CUDA_CALLABLE_MEMBER void imag(value_type __im) { __im_ = __im; }
-
-  CUDA_CALLABLE_MEMBER complex &operator=(float __re) {
-    __re_ = __re;
-    __im_ = value_type();
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator+=(float __re) {
-    __re_ += __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator-=(float __re) {
-    __re_ -= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator*=(float __re) {
-    __re_ *= __re;
-    __im_ *= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator/=(float __re) {
-    __re_ /= __re;
-    __im_ /= __re;
-    return *this;
-  }
-
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator=(const complex<_Xp> &__c) {
-    __re_ = __c.real();
-    __im_ = __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator+=(const complex<_Xp> &__c) {
-    __re_ += __c.real();
-    __im_ += __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator-=(const complex<_Xp> &__c) {
-    __re_ -= __c.real();
-    __im_ -= __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator*=(const complex<_Xp> &__c) {
-    *this = *this * __c;
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator/=(const complex<_Xp> &__c) {
-    *this = *this / __c;
-    return *this;
-  }
-};
-
-template <> class complex<double> {
-  double __re_;
-  double __im_;
-
-public:
-  typedef double value_type;
-
-  /*constexpr*/ CUDA_CALLABLE_MEMBER complex(double __re = 0.0,
-                                             double __im = 0.0)
-      : __re_(__re), __im_(__im) {}
-  /*constexpr*/ complex(const complex<float> &__c);
-
-  /*constexpr*/ CUDA_CALLABLE_MEMBER double real() const { return __re_; }
-  /*constexpr*/ CUDA_CALLABLE_MEMBER double imag() const { return __im_; }
-
-  CUDA_CALLABLE_MEMBER void real(value_type __re) { __re_ = __re; }
-  CUDA_CALLABLE_MEMBER void imag(value_type __im) { __im_ = __im; }
-
-  CUDA_CALLABLE_MEMBER complex &operator=(double __re) {
-    __re_ = __re;
-    __im_ = value_type();
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator+=(double __re) {
-    __re_ += __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator-=(double __re) {
-    __re_ -= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator*=(double __re) {
-    __re_ *= __re;
-    __im_ *= __re;
-    return *this;
-  }
-  CUDA_CALLABLE_MEMBER complex &operator/=(double __re) {
-    __re_ /= __re;
-    __im_ /= __re;
-    return *this;
-  }
-
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator=(const complex<_Xp> &__c) {
-    __re_ = __c.real();
-    __im_ = __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator+=(const complex<_Xp> &__c) {
-    __re_ += __c.real();
-    __im_ += __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator-=(const complex<_Xp> &__c) {
-    __re_ -= __c.real();
-    __im_ -= __c.imag();
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator*=(const complex<_Xp> &__c) {
-    *this = *this * __c;
-    return *this;
-  }
-  template <class _Xp>
-  CUDA_CALLABLE_MEMBER complex &operator/=(const complex<_Xp> &__c) {
-    *this = *this / __c;
-    return *this;
-  }
-};
-
-// constexpr
-inline CUDA_CALLABLE_MEMBER complex<float>::complex(const complex<double> &__c)
-    : __re_(__c.real()), __im_(__c.imag()) {}
-
-// constexpr
-inline CUDA_CALLABLE_MEMBER complex<double>::complex(const complex<float> &__c)
-    : __re_(__c.real()), __im_(__c.imag()) {}
-
-// 26.3.6 operators:
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator+(const complex<_Tp> &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(__x);
-  __t += __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator+(const complex<_Tp> &__x,
-                                                   const _Tp &__y) {
-  complex<_Tp> __t(__x);
-  __t += __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator+(const _Tp &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(__y);
-  __t += __x;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator-(const complex<_Tp> &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(__x);
-  __t -= __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator-(const complex<_Tp> &__x,
-                                                   const _Tp &__y) {
-  complex<_Tp> __t(__x);
-  __t -= __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator-(const _Tp &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(-__y);
-  __t += __x;
-  return __t;
-}
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> operator*(const complex<_Tp> &__z,
-                                            const complex<_Tp> &__w) {
-  _Tp __a = __z.real();
-  _Tp __b = __z.imag();
-  _Tp __c = __w.real();
-  _Tp __d = __w.imag();
-  _Tp __ac = __a * __c;
-  _Tp __bd = __b * __d;
-  _Tp __ad = __a * __d;
-  _Tp __bc = __b * __c;
-  _Tp __x = __ac - __bd;
-  _Tp __y = __ad + __bc;
-  if (isnan(__x) && isnan(__y)) {
-    bool __recalc = false;
-    if (isinf(__a) || isinf(__b)) {
-      __a = copysign(isinf(__a) ? _Tp(1) : _Tp(0), __a);
-      __b = copysign(isinf(__b) ? _Tp(1) : _Tp(0), __b);
-      if (isnan(__c))
-        __c = copysign(_Tp(0), __c);
-      if (isnan(__d))
-        __d = copysign(_Tp(0), __d);
-      __recalc = true;
-    }
-    if (isinf(__c) || isinf(__d)) {
-      __c = copysign(isinf(__c) ? _Tp(1) : _Tp(0), __c);
-      __d = copysign(isinf(__d) ? _Tp(1) : _Tp(0), __d);
-      if (isnan(__a))
-        __a = copysign(_Tp(0), __a);
-      if (isnan(__b))
-        __b = copysign(_Tp(0), __b);
-      __recalc = true;
-    }
-    if (!__recalc &&
-        (isinf(__ac) || isinf(__bd) || isinf(__ad) || isinf(__bc))) {
-      if (isnan(__a))
-        __a = copysign(_Tp(0), __a);
-      if (isnan(__b))
-        __b = copysign(_Tp(0), __b);
-      if (isnan(__c))
-        __c = copysign(_Tp(0), __c);
-      if (isnan(__d))
-        __d = copysign(_Tp(0), __d);
-      __recalc = true;
-    }
-    if (__recalc) {
-      __x = _Tp(INFINITY) * (__a * __c - __b * __d);
-      __y = _Tp(INFINITY) * (__a * __d + __b * __c);
-    }
-  }
-  return complex<_Tp>(__x, __y);
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator*(const complex<_Tp> &__x,
-                                                   const _Tp &__y) {
-  complex<_Tp> __t(__x);
-  __t *= __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator*(const _Tp &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(__y);
-  __t *= __x;
-  return __t;
-}
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> operator/(const complex<_Tp> &__z,
-                                            const complex<_Tp> &__w) {
-  int __ilogbw = 0;
-  _Tp __a = __z.real();
-  _Tp __b = __z.imag();
-  _Tp __c = __w.real();
-  _Tp __d = __w.imag();
-  _Tp __logbw = logb(fmax(fabs(__c), fabs(__d)));
-  if (isfinite(__logbw)) {
-    __ilogbw = static_cast<int>(__logbw);
-    __c = scalbn(__c, -__ilogbw);
-    __d = scalbn(__d, -__ilogbw);
-  }
-  _Tp __denom = __c * __c + __d * __d;
-  _Tp __x = scalbn((__a * __c + __b * __d) / __denom, -__ilogbw);
-  _Tp __y = scalbn((__b * __c - __a * __d) / __denom, -__ilogbw);
-  if (isnan(__x) && isnan(__y)) {
-    if ((__denom == _Tp(0)) && (!isnan(__a) || !isnan(__b))) {
-      __x = copysign(_Tp(INFINITY), __c) * __a;
-      __y = copysign(_Tp(INFINITY), __c) * __b;
-    } else if ((isinf(__a) || isinf(__b)) && isfinite(__c) && isfinite(__d)) {
-      __a = copysign(isinf(__a) ? _Tp(1) : _Tp(0), __a);
-      __b = copysign(isinf(__b) ? _Tp(1) : _Tp(0), __b);
-      __x = _Tp(INFINITY) * (__a * __c + __b * __d);
-      __y = _Tp(INFINITY) * (__b * __c - __a * __d);
-    } else if (isinf(__logbw) && __logbw > _Tp(0) && isfinite(__a) &&
-               isfinite(__b)) {
-      __c = copysign(isinf(__c) ? _Tp(1) : _Tp(0), __c);
-      __d = copysign(isinf(__d) ? _Tp(1) : _Tp(0), __d);
-      __x = _Tp(0) * (__a * __c + __b * __d);
-      __y = _Tp(0) * (__b * __c - __a * __d);
-    }
-  }
-  return complex<_Tp>(__x, __y);
-}
-
-template <>
-CUDA_CALLABLE_MEMBER complex<float> operator/(const complex<float> &__z,
-                                              const complex<float> &__w) {
-  int __ilogbw = 0;
-  float __a = __z.real();
-  float __b = __z.imag();
-  float __c = __w.real();
-  float __d = __w.imag();
-  float __logbw = logbf(fmaxf(fabsf(__c), fabsf(__d)));
-  if (isfinite(__logbw)) {
-    __ilogbw = static_cast<int>(__logbw);
-    __c = scalbnf(__c, -__ilogbw);
-    __d = scalbnf(__d, -__ilogbw);
-  }
-  float __denom = __c * __c + __d * __d;
-  float __x = scalbnf((__a * __c + __b * __d) / __denom, -__ilogbw);
-  float __y = scalbnf((__b * __c - __a * __d) / __denom, -__ilogbw);
-  if (isnan(__x) && isnan(__y)) {
-    if ((__denom == float(0)) && (!isnan(__a) || !isnan(__b))) {
-#pragma warning(suppress : 4756) // Ignore INFINITY related warning
-      __x = copysignf(INFINITY, __c) * __a;
-#pragma warning(suppress : 4756) // Ignore INFINITY related warning
-      __y = copysignf(INFINITY, __c) * __b;
-    } else if ((isinf(__a) || isinf(__b)) && isfinite(__c) && isfinite(__d)) {
-      __a = copysignf(isinf(__a) ? float(1) : float(0), __a);
-      __b = copysignf(isinf(__b) ? float(1) : float(0), __b);
-#pragma warning(suppress : 4756) // Ignore INFINITY related warning
-      __x = INFINITY * (__a * __c + __b * __d);
-#pragma warning(suppress : 4756) // Ignore INFINITY related warning
-      __y = INFINITY * (__b * __c - __a * __d);
-    } else if (isinf(__logbw) && __logbw > float(0) && isfinite(__a) &&
-               isfinite(__b)) {
-      __c = copysignf(isinf(__c) ? float(1) : float(0), __c);
-      __d = copysignf(isinf(__d) ? float(1) : float(0), __d);
-      __x = float(0) * (__a * __c + __b * __d);
-      __y = float(0) * (__b * __c - __a * __d);
-    }
-  }
-  return complex<float>(__x, __y);
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator/(const complex<_Tp> &__x,
-                                                   const _Tp &__y) {
-  return complex<_Tp>(__x.real() / __y, __x.imag() / __y);
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator/(const _Tp &__x,
-                                                   const complex<_Tp> &__y) {
-  complex<_Tp> __t(__x);
-  __t /= __y;
-  return __t;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator+(const complex<_Tp> &__x) {
-  return __x;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> operator-(const complex<_Tp> &__x) {
-  return complex<_Tp>(-__x.real(), -__x.imag());
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator==(const complex<_Tp> &__x,
-                                            const complex<_Tp> &__y) {
-  return __x.real() == __y.real() && __x.imag() == __y.imag();
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator==(const complex<_Tp> &__x,
-                                            const _Tp &__y) {
-  return __x.real() == __y && __x.imag() == 0;
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator==(const _Tp &__x,
-                                            const complex<_Tp> &__y) {
-  return __x == __y.real() && 0 == __y.imag();
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator!=(const complex<_Tp> &__x,
-                                            const complex<_Tp> &__y) {
-  return !(__x == __y);
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator!=(const complex<_Tp> &__x,
-                                            const _Tp &__y) {
-  return !(__x == __y);
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER bool operator!=(const _Tp &__x,
-                                            const complex<_Tp> &__y) {
-  return !(__x == __y);
-}
-
-// 26.3.7 values:
-
-// real
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER _Tp real(const complex<_Tp> &__c) {
-  return __c.real();
-}
-
-inline CUDA_CALLABLE_MEMBER double real(double __re) { return __re; }
-
-inline CUDA_CALLABLE_MEMBER float real(float __re) { return __re; }
-
-// imag
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER _Tp imag(const complex<_Tp> &__c) {
-  return __c.imag();
-}
-
-inline CUDA_CALLABLE_MEMBER double imag(double __re) { return 0; }
-
-inline CUDA_CALLABLE_MEMBER float imag(float __re) { return 0; }
-
-// abs
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER _Tp abs(const complex<_Tp> &__c) {
-  return hypot(__c.real(), __c.imag());
-}
-
-// arg
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER _Tp arg(const complex<_Tp> &__c) {
-  return atan2(__c.imag(), __c.real());
-}
-
-inline CUDA_CALLABLE_MEMBER double arg(double __re) { return atan2(0., __re); }
-
-inline CUDA_CALLABLE_MEMBER float arg(float __re) { return atan2f(0.F, __re); }
-
-// norm
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER _Tp norm(const complex<_Tp> &__c) {
-  if (isinf(__c.real()))
-    return fabs(__c.real());
-  if (isinf(__c.imag()))
-    return fabs(__c.imag());
-  return __c.real() * __c.real() + __c.imag() * __c.imag();
-}
-
-inline CUDA_CALLABLE_MEMBER double norm(double __re) { return __re * __re; }
-
-inline CUDA_CALLABLE_MEMBER float norm(float __re) { return __re * __re; }
-
-// conj
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> conj(const complex<_Tp> &__c) {
-  return complex<_Tp>(__c.real(), -__c.imag());
-}
-
-inline CUDA_CALLABLE_MEMBER complex<double> conj(double __re) {
-  return complex<double>(__re);
-}
-
-inline CUDA_CALLABLE_MEMBER complex<float> conj(float __re) {
-  return complex<float>(__re);
-}
-
-// proj
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> proj(const complex<_Tp> &__c) {
-  complex<_Tp> __r = __c;
-  if (isinf(__c.real()) || isinf(__c.imag()))
-    __r = complex<_Tp>(INFINITY, copysign(_Tp(0), __c.imag()));
-  return __r;
-}
-
-inline CUDA_CALLABLE_MEMBER complex<double> proj(double __re) {
-  if (isinf(__re))
-    __re = fabs(__re);
-  return complex<double>(__re);
-}
-
-inline CUDA_CALLABLE_MEMBER complex<float> proj(float __re) {
-  if (isinf(__re))
-    __re = fabs(__re);
-  return complex<float>(__re);
-}
-
-// polar
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> polar(const _Tp &__rho,
-                                        const _Tp &__theta = _Tp(0)) {
-  if (isnan(__rho) || signbit(__rho))
-    return complex<_Tp>(_Tp(NAN), _Tp(NAN));
-  if (isnan(__theta)) {
-    if (isinf(__rho))
-      return complex<_Tp>(__rho, __theta);
-    return complex<_Tp>(__theta, __theta);
-  }
-  if (isinf(__theta)) {
-    if (isinf(__rho))
-      return complex<_Tp>(__rho, _Tp(NAN));
-    return complex<_Tp>(_Tp(NAN), _Tp(NAN));
-  }
-  _Tp __x = __rho * cos(__theta);
-  if (isnan(__x))
-    __x = 0;
-  _Tp __y = __rho * sin(__theta);
-  if (isnan(__y))
-    __y = 0;
-  return complex<_Tp>(__x, __y);
-}
-
-// log
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> log(const complex<_Tp> &__x) {
-  return complex<_Tp>(log(abs(__x)), arg(__x));
-}
-
-// log10
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> log10(const complex<_Tp> &__x) {
-  return log(__x) / log(_Tp(10));
-}
-
-// sqrt
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> sqrt(const complex<_Tp> &__x) {
-  if (isinf(__x.imag()))
-    return complex<_Tp>(_Tp(INFINITY), __x.imag());
-  if (isinf(__x.real())) {
-    if (__x.real() > _Tp(0))
-      return complex<_Tp>(__x.real(), isnan(__x.imag())
-                                          ? __x.imag()
-                                          : copysign(_Tp(0), __x.imag()));
-    return complex<_Tp>(isnan(__x.imag()) ? __x.imag() : _Tp(0),
-                        copysign(__x.real(), __x.imag()));
-  }
-  return polar(sqrt(abs(__x)), arg(__x) / _Tp(2));
-}
-
-// exp
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> exp(const complex<_Tp> &__x) {
-  _Tp __i = __x.imag();
-  if (isinf(__x.real())) {
-    if (__x.real() < _Tp(0)) {
-      if (!isfinite(__i))
-        __i = _Tp(1);
-    } else if (__i == 0 || !isfinite(__i)) {
-      if (isinf(__i))
-        __i = _Tp(NAN);
-      return complex<_Tp>(__x.real(), __i);
-    }
-  } else if (isnan(__x.real()) && __x.imag() == 0)
-    return __x;
-  _Tp __e = exp(__x.real());
-  return complex<_Tp>(__e * cos(__i), __e * sin(__i));
-}
-
-// pow
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> pow(const complex<_Tp> &__x,
-                                             const complex<_Tp> &__y) {
-  return exp(__y * log(__x));
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> pow(const complex<_Tp> &__x,
-                                             const _Tp &__y) {
-  return pow(__x, complex<_Tp>(__y));
-}
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> pow(const _Tp &__x,
-                                             const complex<_Tp> &__y) {
-  return pow(complex<_Tp>(__x), __y);
-}
-
-// asinh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> asinh(const complex<_Tp> &__x) {
-  const _Tp __pi(atan2(+0., -0.));
-  if (isinf(__x.real())) {
-    if (isnan(__x.imag()))
-      return __x;
-    if (isinf(__x.imag()))
-      return complex<_Tp>(__x.real(), copysign(__pi * _Tp(0.25), __x.imag()));
-    return complex<_Tp>(__x.real(), copysign(_Tp(0), __x.imag()));
-  }
-  if (isnan(__x.real())) {
-    if (isinf(__x.imag()))
-      return complex<_Tp>(__x.imag(), __x.real());
-    if (__x.imag() == 0)
-      return __x;
-    return complex<_Tp>(__x.real(), __x.real());
-  }
-  if (isinf(__x.imag()))
-    return complex<_Tp>(copysign(__x.imag(), __x.real()),
-                        copysign(__pi / _Tp(2), __x.imag()));
-  complex<_Tp> __z = log(__x + sqrt(pow(__x, _Tp(2)) + _Tp(1)));
-  return complex<_Tp>(copysign(__z.real(), __x.real()),
-                      copysign(__z.imag(), __x.imag()));
-}
-
-// acosh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> acosh(const complex<_Tp> &__x) {
-  const _Tp __pi(atan2(+0., -0.));
-  if (isinf(__x.real())) {
-    if (isnan(__x.imag()))
-      return complex<_Tp>(fabs(__x.real()), __x.imag());
-    if (isinf(__x.imag()))
-      if (__x.real() > 0)
-        return complex<_Tp>(__x.real(), copysign(__pi * _Tp(0.25), __x.imag()));
-      else
-        return complex<_Tp>(-__x.real(),
-                            copysign(__pi * _Tp(0.75), __x.imag()));
-    if (__x.real() < 0)
-      return complex<_Tp>(-__x.real(), copysign(__pi, __x.imag()));
-    return complex<_Tp>(__x.real(), copysign(_Tp(0), __x.imag()));
-  }
-  if (isnan(__x.real())) {
-    if (isinf(__x.imag()))
-      return complex<_Tp>(fabs(__x.imag()), __x.real());
-    return complex<_Tp>(__x.real(), __x.real());
-  }
-  if (isinf(__x.imag()))
-    return complex<_Tp>(fabs(__x.imag()), copysign(__pi / _Tp(2), __x.imag()));
-  complex<_Tp> __z = log(__x + sqrt(pow(__x, _Tp(2)) - _Tp(1)));
-  return complex<_Tp>(copysign(__z.real(), _Tp(0)),
-                      copysign(__z.imag(), __x.imag()));
-}
-
-// atanh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> atanh(const complex<_Tp> &__x) {
-  const _Tp __pi(atan2(+0., -0.));
-  if (isinf(__x.imag())) {
-    return complex<_Tp>(copysign(_Tp(0), __x.real()),
-                        copysign(__pi / _Tp(2), __x.imag()));
-  }
-  if (isnan(__x.imag())) {
-    if (isinf(__x.real()) || __x.real() == 0)
-      return complex<_Tp>(copysign(_Tp(0), __x.real()), __x.imag());
-    return complex<_Tp>(__x.imag(), __x.imag());
-  }
-  if (isnan(__x.real())) {
-    return complex<_Tp>(__x.real(), __x.real());
-  }
-  if (isinf(__x.real())) {
-    return complex<_Tp>(copysign(_Tp(0), __x.real()),
-                        copysign(__pi / _Tp(2), __x.imag()));
-  }
-  if (fabs(__x.real()) == _Tp(1) && __x.imag() == _Tp(0)) {
-    return complex<_Tp>(copysign(_Tp(INFINITY), __x.real()),
-                        copysign(_Tp(0), __x.imag()));
-  }
-  complex<_Tp> __z = log((_Tp(1) + __x) / (_Tp(1) - __x)) / _Tp(2);
-  return complex<_Tp>(copysign(__z.real(), __x.real()),
-                      copysign(__z.imag(), __x.imag()));
-}
-
-// sinh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> sinh(const complex<_Tp> &__x) {
-  if (isinf(__x.real()) && !isfinite(__x.imag()))
-    return complex<_Tp>(__x.real(), _Tp(NAN));
-  if (__x.real() == 0 && !isfinite(__x.imag()))
-    return complex<_Tp>(__x.real(), _Tp(NAN));
-  if (__x.imag() == 0 && !isfinite(__x.real()))
-    return __x;
-  return complex<_Tp>(sinh(__x.real()) * cos(__x.imag()),
-                      cosh(__x.real()) * sin(__x.imag()));
-}
-
-// cosh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> cosh(const complex<_Tp> &__x) {
-  if (isinf(__x.real()) && !isfinite(__x.imag()))
-    return complex<_Tp>(fabs(__x.real()), _Tp(NAN));
-  if (__x.real() == 0 && !isfinite(__x.imag()))
-    return complex<_Tp>(_Tp(NAN), __x.real());
-  if (__x.real() == 0 && __x.imag() == 0)
-    return complex<_Tp>(_Tp(1), __x.imag());
-  if (__x.imag() == 0 && !isfinite(__x.real()))
-    return complex<_Tp>(fabs(__x.real()), __x.imag());
-  return complex<_Tp>(cosh(__x.real()) * cos(__x.imag()),
-                      sinh(__x.real()) * sin(__x.imag()));
-}
-
-// tanh
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> tanh(const complex<_Tp> &__x) {
-  if (isinf(__x.real())) {
-    if (!isfinite(__x.imag()))
-      return complex<_Tp>(_Tp(1), _Tp(0));
-    return complex<_Tp>(_Tp(1), copysign(_Tp(0), sin(_Tp(2) * __x.imag())));
-  }
-  if (isnan(__x.real()) && __x.imag() == 0)
-    return __x;
-  _Tp __2r(_Tp(2) * __x.real());
-  _Tp __2i(_Tp(2) * __x.imag());
-  _Tp __d(cosh(__2r) + cos(__2i));
-  return complex<_Tp>(sinh(__2r) / __d, sin(__2i) / __d);
-}
-
-// asin
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> asin(const complex<_Tp> &__x) {
-  complex<_Tp> __z = asinh(complex<_Tp>(-__x.imag(), __x.real()));
-  return complex<_Tp>(__z.imag(), -__z.real());
-}
-
-// acos
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> acos(const complex<_Tp> &__x) {
-  const _Tp __pi(atan2(+0., -0.));
-  if (isinf(__x.real())) {
-    if (isnan(__x.imag()))
-      return complex<_Tp>(__x.imag(), __x.real());
-    if (isinf(__x.imag())) {
-      if (__x.real() < _Tp(0))
-        return complex<_Tp>(_Tp(0.75) * __pi, -__x.imag());
-      return complex<_Tp>(_Tp(0.25) * __pi, -__x.imag());
-    }
-    if (__x.real() < _Tp(0))
-      return complex<_Tp>(__pi, signbit(__x.imag()) ? -__x.real() : __x.real());
-    return complex<_Tp>(_Tp(0), signbit(__x.imag()) ? __x.real() : -__x.real());
-  }
-  if (isnan(__x.real())) {
-    if (isinf(__x.imag()))
-      return complex<_Tp>(__x.real(), -__x.imag());
-    return complex<_Tp>(__x.real(), __x.real());
-  }
-  if (isinf(__x.imag()))
-    return complex<_Tp>(__pi / _Tp(2), -__x.imag());
-  if (__x.real() == 0)
-    return complex<_Tp>(__pi / _Tp(2), -__x.imag());
-  complex<_Tp> __z = log(__x + sqrt(pow(__x, _Tp(2)) - _Tp(1)));
-  if (signbit(__x.imag()))
-    return complex<_Tp>(fabs(__z.imag()), fabs(__z.real()));
-  return complex<_Tp>(fabs(__z.imag()), -fabs(__z.real()));
-}
-
-// atan
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> atan(const complex<_Tp> &__x) {
-  complex<_Tp> __z = atanh(complex<_Tp>(-__x.imag(), __x.real()));
-  return complex<_Tp>(__z.imag(), -__z.real());
-}
-
-// sin
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> sin(const complex<_Tp> &__x) {
-  complex<_Tp> __z = sinh(complex<_Tp>(-__x.imag(), __x.real()));
-  return complex<_Tp>(__z.imag(), -__z.real());
-}
-
-// cos
-
-template <class _Tp>
-inline CUDA_CALLABLE_MEMBER complex<_Tp> cos(const complex<_Tp> &__x) {
-  return cosh(complex<_Tp>(-__x.imag(), __x.real()));
-}
-
-// tan
-
-template <class _Tp>
-CUDA_CALLABLE_MEMBER complex<_Tp> tan(const complex<_Tp> &__x) {
-  complex<_Tp> __z = tanh(complex<_Tp>(-__x.imag(), __x.real()));
-  return complex<_Tp>(__z.imag(), -__z.real());
-}
-
-template <class _Tp, class _CharT, class _Traits>
-std::basic_istream<_CharT, _Traits> &
-operator>>(std::basic_istream<_CharT, _Traits> &__is, complex<_Tp> &__x) {
-  if (__is.good()) {
-    ws(__is);
-    if (__is.peek() == _CharT('(')) {
-      __is.get();
-      _Tp __r;
-      __is >> __r;
-      if (!__is.fail()) {
-        ws(__is);
-        _CharT __c = __is.peek();
-        if (__c == _CharT(',')) {
-          __is.get();
-          _Tp __i;
-          __is >> __i;
-          if (!__is.fail()) {
-            ws(__is);
-            __c = __is.peek();
-            if (__c == _CharT(')')) {
-              __is.get();
-              __x = complex<_Tp>(__r, __i);
-            } else
-              __is.setstate(std::ios_base::failbit);
-          } else
-            __is.setstate(std::ios_base::failbit);
-        } else if (__c == _CharT(')')) {
-          __is.get();
-          __x = complex<_Tp>(__r, _Tp(0));
-        } else
-          __is.setstate(std::ios_base::failbit);
-      } else
-        __is.setstate(std::ios_base::failbit);
-    } else {
-      _Tp __r;
-      __is >> __r;
-      if (!__is.fail())
-        __x = complex<_Tp>(__r, _Tp(0));
-      else
-        __is.setstate(std::ios_base::failbit);
-    }
-  } else
-    __is.setstate(std::ios_base::failbit);
-  return __is;
-}
-
-template <class _Tp, class _CharT, class _Traits>
-std::basic_ostream<_CharT, _Traits> &
-operator<<(std::basic_ostream<_CharT, _Traits> &__os, const complex<_Tp> &__x) {
-  std::basic_ostringstream<_CharT, _Traits> __s;
-  __s.flags(__os.flags());
-  __s.imbue(__os.getloc());
-  __s.precision(__os.precision());
-  __s << '(' << __x.real() << ',' << __x.imag() << ')';
-  return __os << __s.str();
-}
-
-//} // close namespace cuda_complex
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator*(const complex<U> &complexNumber,
-                                    const V &scalar) -> complex<U> {
-  return complex<U>(real(complexNumber) * scalar, imag(complexNumber) * scalar);
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator*(const V &scalar,
-                                    const complex<U> &complexNumber)
-    -> complex<U> {
-  return complex<U>(real(complexNumber) * scalar, imag(complexNumber) * scalar);
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator+(const complex<U> &complexNumber,
-                                    const V &scalar) -> complex<U> {
-  return complex<U>(real(complexNumber) + scalar, imag(complexNumber));
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator+(const V &scalar,
-                                    const complex<U> &complexNumber)
-    -> complex<U> {
-  return complex<U>(real(complexNumber) + scalar, imag(complexNumber));
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator-(const complex<U> &complexNumber,
-                                    const V &scalar) -> complex<U> {
-  return complex<U>(real(complexNumber) - scalar, imag(complexNumber));
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator-(const V &scalar,
-                                    const complex<U> &complexNumber)
-    -> complex<U> {
-  return complex<U>(scalar - real(complexNumber), imag(complexNumber));
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator/(const complex<U> &complexNumber,
-                                    const V scalar) -> complex<U> {
-  return complex<U>(real(complexNumber) / scalar, imag(complexNumber) / scalar);
-}
-
-template <class U, class V>
-CUDA_CALLABLE_MEMBER auto operator/(const V scalar,
-                                    const complex<U> &complexNumber)
-    -> complex<U> {
-  return complex<U>(scalar, 0) / complexNumber;
-}
-
-using ComplexDouble = complex<double>;
-using ComplexFloat = complex<float>;
-#endif // CUDA_COMPLEX_HPP
-}