numba_dpex.core.utils.call_kernel_builder¶

Module that contains numba style wrapper around sycl kernel submit.

Overview¶

Classes¶
`SPIRVKernelModule`	Represents SPIRV binary code and function name in this binary
`KernelLaunchIRBuilder`	Helper class to build the LLVM IR for the submission of a kernel.

Function¶
`get_queue_from_llvm_values`(ctx, builder, ty_kernel_args, ll_kernel_args)	Get the sycl queue from the first USMNdArray argument. Prior passes

Attributes¶
`MAX_SIZE_OF_SYCL_RANGE`	-
`OPEN_CL_OPT_DISABLE_FLAG`	-
`L0_OPT_DISABLE_FLAG`	-

Classes¶

class SPIRVKernelModule¶

Bases: NamedTuple

Represents SPIRV binary code and function name in this binary

Overview

Attributes¶
`kernel_name`	-
`kernel_bitcode`	-

Members

kernel_name: str¶

kernel_bitcode: bytes¶

class KernelLaunchIRBuilder(context: numba.core.cpu.CPUContext, builder: llvmlite.ir.builder.IRBuilder, kernel_dmm: numba.core.datamodel.DataModelManager)¶

Helper class to build the LLVM IR for the submission of a kernel.

The class generates LLVM IR inside the current LLVM module that is needed for submitting kernels. The LLVM Values that

Overview

Methods¶
`dpexrt`()	Dpex runtime context.
`get_queue`(exec_queue)	Allocates memory on the stack to store a DPCTLSyclQueueRef.
`set_kernel`(sycl_kernel_ref)	Sets kernel to the argument list.
`set_kernel_from_spirv`(kernel_module, debug)	Sets kernel to the argument list from the SPIRV bytecode.
`set_queue`(sycl_queue_ref)	Sets queue to the argument list.
`set_queue_from_arguments`()	Sets the sycl queue from the first USMNdArray argument provided
`set_range`(global_range, local_range)	Sets global and local range if provided to the argument list.
`set_range_from_indexer`(ty_indexer_arg, ll_index_arg)	Returns two lists of LLVM IR Values that hold the unboxed extents of
`set_arguments`(ty_kernel_args, kernel_args)	Sets flattened kernel args, kernel arg types and number of those
`set_arguments_form_tuple`(ty_kernel_args_tuple, ll_kernel_args_tuple)	Sets flattened kernel args, kernel arg types and number of those
`set_dependent_events`(dep_events)	Sets dependent events to the argument list.
`set_dependent_events_from_tuple`(ty_dependent_events, ll_dependent_events)	Set’s dependent events from tuple represented by LLVM IR.
`submit`()	Submits kernel by calling sycl.dpctl_queue_submit_range or
`acquire_meminfo_and_submit_release`()	Schedule sycl host task to release nrt meminfo of the arguments used

Members

dpexrt()¶: Dpex runtime context.

get_queue(exec_queue: dpctl.SyclQueue) → llvmlite.ir.Instruction¶

Allocates memory on the stack to store a DPCTLSyclQueueRef.

Returns: A LLVM Value storing the pointer to the SYCL queue created using the filter string for the Python exec_queue (dpctl.SyclQueue).

set_kernel(sycl_kernel_ref: llvmlite.ir.Instruction)¶: Sets kernel to the argument list.

set_kernel_from_spirv(kernel_module: SPIRVKernelModule, debug=False)¶

Sets kernel to the argument list from the SPIRV bytecode.

It pastes bytecode as a constant string and create kernel bundle from it using SYCL API. It caches kernel, so it won’t be sent to device second time.

set_queue(sycl_queue_ref: llvmlite.ir.Instruction)¶: Sets queue to the argument list.

set_queue_from_arguments()¶: Sets the sycl queue from the first USMNdArray argument provided earlier.

set_range(global_range: list, local_range: list = None)¶: Sets global and local range if provided to the argument list.

set_range_from_indexer(ty_indexer_arg: RangeType | NdRangeType, ll_index_arg: llvmlite.ir.BaseStructType)¶: Returns two lists of LLVM IR Values that hold the unboxed extents of a Python Range or NdRange object.

set_arguments(ty_kernel_args: list[numba.core.types.Type], kernel_args: list[llvmlite.ir.Instruction])¶: Sets flattened kernel args, kernel arg types and number of those arguments to the argument list.

set_arguments_form_tuple(ty_kernel_args_tuple: numba.core.types.containers.UniTuple, ll_kernel_args_tuple: llvmlite.ir.Instruction)¶: Sets flattened kernel args, kernel arg types and number of those arguments to the argument list based on the arguments stored in tuple.

set_dependent_events(dep_events: list[llvmlite.ir.Instruction])¶: Sets dependent events to the argument list.

set_dependent_events_from_tuple(ty_dependent_events: numba.core.types.containers.UniTuple, ll_dependent_events: llvmlite.ir.Instruction)¶

Set’s dependent events from tuple represented by LLVM IR.

Parameters:: ll_dependent_events – tuple of numba’s data models.

submit() → llvmlite.ir.Instruction¶: Submits kernel by calling sycl.dpctl_queue_submit_range or sycl.dpctl_queue_submit_ndrange. Must be called after all arguments set.

acquire_meminfo_and_submit_release() → llvmlite.ir.Instruction¶: Schedule sycl host task to release nrt meminfo of the arguments used to run job. Use it to keep arguments alive during kernel execution.

Functions¶

get_queue_from_llvm_values(ctx: numba.core.cpu.CPUContext, builder: llvmlite.ir.builder.IRBuilder, ty_kernel_args: list[numba.core.types.Type], ll_kernel_args: list[llvmlite.ir.Instruction])¶: Get the sycl queue from the first USMNdArray argument. Prior passes before lowering make sure that compute-follows-data is enforceable for a specific call to a kernel. As such, at the stage of lowering the queue from the first USMNdArray argument can be extracted.

Attributes¶

MAX_SIZE_OF_SYCL_RANGE = 3¶

OPEN_CL_OPT_DISABLE_FLAG = '-cl-opt-disable'¶

L0_OPT_DISABLE_FLAG = '-g'¶