numba_vulkan.target#

The Vulkan target, registered with Numba’s target extension API.

Attributes#

`TARGET_NAME`
`exception_table`
`vulkan_target`

Classes#

`Vulkan`	Marks the Vulkan target in Numba's target hierarchy.
`VulkanTypingContext`	Typing context: which functions and operations type-check on Vulkan.
`VulkanCallConv`	Calling convention of functions compiled for Vulkan.
`VulkanTargetContext`	Target context: how typed operations are lowered to LLVM IR.
`VulkanTargetOptions`	Options accepted by the target; none beyond Numba's defaults.
`VulkanTarget`	Target descriptor: gives access to the typing and target contexts.

Module Contents#

numba_vulkan.target.TARGET_NAME = 'vulkan'#

class numba_vulkan.target.Vulkan#

Marks the Vulkan target in Numba’s target hierarchy.

It derives from Numba’s generic GPU target, so overloads registered for target="gpu" or target="generic" apply to it as well.

class numba_vulkan.target.VulkanTypingContext#

Typing context: which functions and operations type-check on Vulkan.

load_additional_registries()#

Install the typing registries beyond Numba’s built-in one.

These are the Vulkan-specific declarations and Numba’s math, cmath, NumPy and enum declarations.

resolve_value_type(val)#

Type a Python value used as a global or closure variable.

A function compiled for the CPU (numba.njit) is replaced by a Vulkan dispatcher for the same Python function, so that calling it from a kernel recompiles it for Vulkan.

Parameters:: val (object) – The value.
Return type:: numba.types.Type

numba_vulkan.target.exception_table#

class numba_vulkan.target.VulkanCallConv#

Calling convention of functions compiled for Vulkan.

Numba’s minimal convention: the return value is written through a pointer argument and the function returns a status code. It never reaches the shader, because everything is inlined; the status of the kernel itself is stored for the host by the entry point (see numba_vulkan.compiler.compile_kernel).

class numba_vulkan.target.VulkanTargetContext(typingctx, target=TARGET_NAME)#

Target context: how typed operations are lowered to LLVM IR.

Parameters:

typingctx (VulkanTypingContext) – The typing context.
target (str, optional) – Name of the target in Numba’s registry.

Variables:

narrow_math (bool) – Whether float64 transcendental functions are evaluated in float32 when libclc is unavailable. Set per compiled function through subtarget.
fast_math (bool) – Whether float32 math uses the device’s built-in functions instead of libclc. Set per compiled function through subtarget.

implement_powi_as_math_call = True#

strict_alignment = True#

narrow_math = False#

fast_math = False#

data_model_manager#

create_module(name)#

Create an empty LLVM module for this target.

Parameters:: name (str) – Module name.
Return type:: llvmlite.ir.Module

init()#: Create the code generator. Called once by Numba’s base context.

load_additional_registries()#

Install the lowering registries beyond Numba’s built-in one.

Importing Numba’s implementation modules fills its built-in registry with number, tuple, range and array metadata operations. The Vulkan registries add array element access, global_id and math.

get_function(fn, sig, _firstcall=True)#

Find the implementation of a function for a signature.

Parameters:

fn (callable or numba.types.Type) – The function.
sig (numba.core.typing.Signature) – Signature of the call.
_firstcall (bool, optional) – Used internally by Numba.

Returns:

impl(builder, args) emitting the call.

Return type:

callable

Notes

Integer-exponent powers are intercepted here; see numba_vulkan.mathimpl.power_override.

get_ufunc_info(ufunc_key)#

Look up how a NumPy ufunc is lowered.

Parameters:: ufunc_key (numpy.ufunc) – The ufunc.
Returns:: Loop signature to implementation; see numba_vulkan.ufuncs.
Return type:: dict
Raises:: KeyError – If the ufunc is not supported on Vulkan.

make_constant_array(builder, aryty, arr)#

Build the value of a NumPy array used as a global constant.

The data is not emitted into the shader. The array’s type names a buffer that the runtime fills once per kernel (see numba_vulkan.buffers.constant_binding); the value built here is the metadata that goes with it.

Parameters:

builder (llvmlite.ir.IRBuilder) – Builder positioned where the code is emitted.
aryty (VulkanArray) – Type of the array.
arr (numpy.ndarray) – The array.

Return type:

llvmlite.ir.Value

codegen()#

The code generator of this context.

Return type:: numba_vulkan.codegen.VulkanCodegen

property target_data#

LLVM data layout of the SPIR-V target, used for type sizes.

Return type:: llvmlite.binding.TargetData

property call_conv#

The calling convention of compiled functions.

Return type:: VulkanCallConv

mangler(name, argtypes, *, abi_tags=(), uid=None)#

Mangle a function name with its argument types.

Parameters:

name (str) – Qualified function name.
argtypes (sequence of numba.types.Type) – Argument types.
abi_tags (sequence of str, optional) – ABI tags to embed.
uid (int, optional) – Unique identifier of the function.

Returns:

The symbol name, in Itanium C++ ABI style.

Return type:

str

class numba_vulkan.target.VulkanTargetOptions#: Options accepted by the target; none beyond Numba’s defaults.

class numba_vulkan.target.VulkanTarget(name)#

Target descriptor: gives access to the typing and target contexts.

The contexts are created on first use, so importing the package does not initialise LLVM.

Parameters:: name (str) – Name of the target in Numba’s registry.

options#

property typing_context#

The shared typing context.

Return type:: VulkanTypingContext

property target_context#

The shared target context.

Return type:: VulkanTargetContext

numba_vulkan.target.vulkan_target#