Code Generation

Invocation

create_kernel

Create a kernel function from a set of assignments.

Configuration

The code generation driver (create_kernel, but also DefaultKernelCreationDriver) can be configured by passing it a CreateKernelConfig object. This object can be constructed incrementally:

cfg = ps.CreateKernelConfig()
cfg.default_dtype = "float32"
cfg.target = ps.Target.X86_AVX
cfg.cpu.openmp.enable = True
cfg.cpu.vectorize.enable = True
cfg.cpu.vectorize.assume_inner_stride_one = True

Options and Option Categories

The following options and option categories are exposed by the configuration object:

Target Specification

target

The code generation target.

Data Types

default_dtype	Default numeric data type.
index_dtype	Data type used for all index calculations.

Iteration Space

ghost_layers	Specifies the number of ghost layers of the iteration region.
iteration_slice	Specifies the kernel's iteration slice.
index_field	Index field for a sparse kernel.

Kernel Constraint Checks

allow_double_writes	If True, don't check if every field is only written at a single location.
skip_independence_check	By default the assignment list is checked for read/write independence.

Target-Specific Options

The following categories with target-specific options are exposed:

cpu	Options for CPU kernels
gpu	Options for GPU kernels
sycl	Options for SYCL kernels

Kernel Object and Just-In-Time Compilation

function_name	Name of the generated function
jit	Just-in-time compiler used to compile and load the kernel for invocation from the current Python environment.

Configuration System Classes

CreateKernelConfig	Options for create_kernel.
CpuOptions	Configuration options specific to CPU targets.
OpenMpOptions	Configuration options controlling automatic OpenMP instrumentation.
VectorizationOptions	Configuration for the auto-vectorizer.
GpuOptions	Configuration options specific to GPU targets.
SyclOptions	Options specific to the SYCL target.
GpuIndexingScheme	Available index translation schemes for GPU kernels.

AUTO	Special value that can be passed to some options for invoking automatic behaviour.

Implementation Details

Option	Option descriptor.
BasicOption	Subclass of Option where Arg_T == Option_T.
Category	Descriptor for a category of options.
ConfigBase	Base class for configuration categories.

Target Specification

Target

The Target enumeration represents all possible targets that can be used for code generation.

Code Generation Driver and Helper Classes

Driver

DefaultKernelCreationDriver	Drives the default kernel creation sequence.
KernelFactory	Factory for wrapping up backend and IR objects into exportable kernels and function objects.

get_driver

Create a code generation driver object from the given configuration.

CPU Loop Strategies

DefaultCpuLoopStrategies

GPU Indexing

HardwareProperties
GpuIndexMappingStrategy
GpuIndexing	Factory for GPU indexing objects required during code generation.
GpuLaunchConfiguration	Base class for launch configurations for CUDA and HIP kernels.
AutomaticLaunchConfiguration	Launch configuration that is dynamically computed from kernel parameters.
ManualLaunchConfiguration	Manual GPU launch configuration.
DynamicBlockSizeLaunchConfiguration	GPU launch configuration that dynamically computes the grid size from either the default block size or a computed block size.

Output Code Objects

Kernel	A pystencils kernel.
GpuKernel	Internal representation of a kernel function targeted at CUDA GPUs.
Parameter	Parameter to an output object of the code generator.
Lambda	A one-line function emitted by the code generator as an auxiliary object.