Beaver.MLIR.Transform (beaver v0.4.8)

Copy Markdown

Transform dialect execution and transformations MLIR provides by default.

apply_named_sequence/3 executes a named Transform dialect sequence without routing it through a pass manager. Transform IR may be supplied as a module, an operation nested in a module, textual MLIR, bytecode, or a resolved tuning schedule produced by Beaver.MLIR.Transform.Schedule.

Handles remain owned by MLIR's transform state and never escape this call. Keeping expensive checks enabled (the default) makes use-after-consume handle errors visible as processed diagnostics.

Summary

Types

execution_option()
schedule_input()

Functions

any_op_type(opts \\ [])

Creates Transform dialect's !transform.any_op type.

any_param_type(opts \\ [])

Creates Transform dialect's !transform.any_param type.

any_value_type(opts \\ [])

Creates Transform dialect's !transform.any_value type.

apply_named_sequence(payload, schedule, opts \\ [])

Applies a named Transform dialect sequence to a payload operation or module.

apply_named_sequence!(payload, schedule, opts \\ [])

Bang variant of apply_named_sequence/3.

bubble_down_memory_space_casts()

Bubbles down memory-space cast operations.

bubble_down_memory_space_casts(composer_or_op)
canonicalize()

Canonicalize operations

canonicalize(composer_or_op)
composite_fixed_point_pass()

Composite fixed point pass

composite_fixed_point_pass(composer_or_op)
control_flow_sink()

Sink operations into conditional blocks

control_flow_sink(composer_or_op)
convert_elementwise_to_linalg()

Convert ElementwiseMappable ops to linalg

convert_elementwise_to_linalg(composer_or_op)
convert_linalg_to_affine_loops()

Lower the operations from the linalg dialect into affine loops

convert_linalg_to_affine_loops(composer_or_op)
convert_linalg_to_loops()

Lower the operations from the linalg dialect into loops

convert_linalg_to_loops(composer_or_op)
convert_linalg_to_parallel_loops()

Lower the operations from the linalg dialect into parallel loops

convert_linalg_to_parallel_loops(composer_or_op)
cse()

Eliminate common sub-expressions

cse(composer_or_op)
execute(payload, schedule, opts \\ [])

Alias for apply_named_sequence/3.

execute!(payload, schedule, opts \\ [])

Alias for apply_named_sequence!/3.

generate_runtime_verification()

Generate additional runtime op verification checks

generate_runtime_verification(composer_or_op)
gpu_async_region()

Make GPU ops async

gpu_async_region(composer_or_op)
gpu_decompose_memrefs()

Decomposes memref index computation into explicit ops.

gpu_decompose_memrefs(composer_or_op)
gpu_eliminate_barriers()

Erase unnecessary barriers

gpu_eliminate_barriers(composer_or_op)
gpu_kernel_outlining()

Outline gpu.launch bodies to kernel functions

gpu_kernel_outlining(composer_or_op)
gpu_launch_sink_index_computations()

Sink index computations into gpu.launch body

gpu_launch_sink_index_computations(composer_or_op)
gpu_map_parallel_loops()

Greedily maps loops to GPU hardware dimensions.

gpu_map_parallel_loops(composer_or_op)
gpu_module_to_binary()

Transforms a GPU module into a GPU binary.

gpu_module_to_binary(composer_or_op)
inline()

Inline function calls

inline(composer_or_op)
linalg_block_pack_matmul()

Convert linalg matmul ops to block layout and back

linalg_block_pack_matmul(composer_or_op)
linalg_fold_into_elementwise()

Fold transpose and broadcast ops into elementwise consumers

linalg_fold_into_elementwise(composer_or_op)
linalg_fold_unit_extent_dims()

Remove unit-extent dimension in Linalg ops on tensors

linalg_fold_unit_extent_dims(composer_or_op)
linalg_fuse_elementwise_ops()

Fuse elementwise operations on tensors

linalg_fuse_elementwise_ops(composer_or_op)
linalg_generalize_named_ops()

Convert named ops into generic ops

linalg_generalize_named_ops(composer_or_op)
linalg_inline_scalar_operands()

Inline scalar operands into linalg generic ops

linalg_inline_scalar_operands(composer_or_op)
linalg_morph_ops()

Convert linalg ops between forms

linalg_morph_ops(composer_or_op)
linalg_specialize_generic_ops()

Convert generic ops back to named ops

linalg_specialize_generic_ops(composer_or_op)
loop_invariant_code_motion()

Hoist loop invariant instructions outside of the loop

loop_invariant_code_motion(composer_or_op)
loop_invariant_subset_hoisting()

Hoist loop invariant subset ops outside of the loop

loop_invariant_subset_hoisting(composer_or_op)
mem2reg()

Promotes memory slots into values.

mem2reg(composer_or_op)
nvvm_attach_target()

Attaches an NVVM target attribute to a GPU Module.

nvvm_attach_target(composer_or_op)
operation_type(operation_name, opts \\ [])

Creates an operation-specific Transform handle type.

param_type(type)

Creates a Transform parameter type wrapping an MLIR type.

print_ir()

Print IR on the debug stream

print_ir(composer_or_op)
print_op_stats()

Print statistics of operations

print_op_stats(composer_or_op)
remove_dead_values()

Remove dead values

remove_dead_values(composer_or_op)
rocdl_attach_target()

Attaches a ROCDL target attribute to a GPU Module.

rocdl_attach_target(composer_or_op)
sccp()

Sparse Conditional Constant Propagation

sccp(composer_or_op)
simplify_depthwise_conv()

Simplify depthwise convolution.

simplify_depthwise_conv(composer_or_op)
snapshot_op_locations()

Generate new locations from the current IR

snapshot_op_locations(composer_or_op)
spirv_attach_target()

Attaches an SPIR-V target attribute to a GPU Module.

spirv_attach_target(composer_or_op)
sroa()

Scalar Replacement of Aggregates

sroa(composer_or_op)
strip_debuginfo()

Strip debug info from all operations

strip_debuginfo(composer_or_op)
symbol_dce()

Eliminate dead symbols

symbol_dce(composer_or_op)
symbol_privatize()

Mark symbols private

symbol_privatize(composer_or_op)
topological_sort()

Sort regions without SSA dominance in topological order

topological_sort(composer_or_op)
trivial_dce()

Remove trivially dead operations and blocks

trivial_dce(composer_or_op)
view_op_graph()

Print Graphviz visualization of an operation

view_op_graph(composer_or_op)
xevm_attach_target()

Attaches a XeVM target attribute to a GPU Module.

xevm_attach_target(composer_or_op)

Types

execution_option()

@type execution_option() ::
  {:sequence, String.t()}
  | {:expensive_checks, boolean()}
  | {:enforce_single_top_level_transform_op, boolean()}

schedule_input()

@type schedule_input() ::
  Beaver.MLIR.Module.t()
  | Beaver.MLIR.Operation.t()
  | binary()
  | Beaver.MLIR.Transform.Schedule.Resolved.t()

Functions

any_op_type(opts \\ [])

Creates Transform dialect's !transform.any_op type.

any_param_type(opts \\ [])

Creates Transform dialect's !transform.any_param type.

any_value_type(opts \\ [])

Creates Transform dialect's !transform.any_value type.

apply_named_sequence(payload, schedule, opts \\ [])

@spec apply_named_sequence(
  Beaver.MLIR.Module.t() | Beaver.MLIR.Operation.t(),
  schedule_input(),
  [execution_option()]
) ::
  {:ok, Beaver.MLIR.Transform.Result.t()}
  | {:error, Beaver.MLIR.Transform.Error.t()}

Applies a named Transform dialect sequence to a payload operation or module.

The default sequence is __transform_main. LLVM's expensive handle checks and single-top-level-transform enforcement both default to true and can be configured independently.

apply_named_sequence!(payload, schedule, opts \\ [])

@spec apply_named_sequence!(
  Beaver.MLIR.Module.t() | Beaver.MLIR.Operation.t(),
  schedule_input(),
  [execution_option()]
) :: Beaver.MLIR.Module.t() | Beaver.MLIR.Operation.t()

Bang variant of apply_named_sequence/3.

bubble_down_memory_space_casts()

Bubbles down memory-space cast operations.

Argument name in MLIR CLI

bubble-down-memory-space-casts

Pass name in TableGen

BubbleDownMemorySpaceCasts

bubble_down_memory_space_casts(composer_or_op)

canonicalize()

Canonicalize operations

Argument name in MLIR CLI

canonicalize

Pass name in TableGen

CanonicalizerPass

canonicalize(composer_or_op)

composite_fixed_point_pass()

Composite fixed point pass

Argument name in MLIR CLI

composite-fixed-point-pass

Pass name in TableGen

CompositeFixedPointPass

composite_fixed_point_pass(composer_or_op)

control_flow_sink()

Sink operations into conditional blocks

Argument name in MLIR CLI

control-flow-sink

Pass name in TableGen

ControlFlowSinkPass

control_flow_sink(composer_or_op)

convert_elementwise_to_linalg()

Convert ElementwiseMappable ops to linalg

Argument name in MLIR CLI

convert-elementwise-to-linalg

Pass name in TableGen

ConvertElementwiseToLinalgPass

convert_elementwise_to_linalg(composer_or_op)

convert_linalg_to_affine_loops()

Lower the operations from the linalg dialect into affine loops

Argument name in MLIR CLI

convert-linalg-to-affine-loops

Pass name in TableGen

ConvertLinalgToAffineLoopsPass

convert_linalg_to_affine_loops(composer_or_op)

convert_linalg_to_loops()

Lower the operations from the linalg dialect into loops

Argument name in MLIR CLI

convert-linalg-to-loops

Pass name in TableGen

ConvertLinalgToLoopsPass

convert_linalg_to_loops(composer_or_op)

convert_linalg_to_parallel_loops()

Lower the operations from the linalg dialect into parallel loops

Argument name in MLIR CLI

convert-linalg-to-parallel-loops

Pass name in TableGen

ConvertLinalgToParallelLoopsPass

convert_linalg_to_parallel_loops(composer_or_op)

cse()

Eliminate common sub-expressions

Argument name in MLIR CLI

cse

Pass name in TableGen

CSEPass

cse(composer_or_op)

execute(payload, schedule, opts \\ [])

Alias for apply_named_sequence/3.

execute!(payload, schedule, opts \\ [])

Alias for apply_named_sequence!/3.

generate_runtime_verification()

Generate additional runtime op verification checks

Argument name in MLIR CLI

generate-runtime-verification

Pass name in TableGen

GenerateRuntimeVerificationPass

generate_runtime_verification(composer_or_op)

gpu_async_region()

Make GPU ops async

Argument name in MLIR CLI

gpu-async-region

Pass name in TableGen

GpuAsyncRegionPass

gpu_async_region(composer_or_op)

gpu_decompose_memrefs()

Decomposes memref index computation into explicit ops.

Argument name in MLIR CLI

gpu-decompose-memrefs

Pass name in TableGen

GpuDecomposeMemrefsPass

gpu_decompose_memrefs(composer_or_op)

gpu_eliminate_barriers()

Erase unnecessary barriers

Argument name in MLIR CLI

gpu-eliminate-barriers

Pass name in TableGen

GpuEliminateBarriers

gpu_eliminate_barriers(composer_or_op)

gpu_kernel_outlining()

Outline gpu.launch bodies to kernel functions

Argument name in MLIR CLI

gpu-kernel-outlining

Pass name in TableGen

GpuKernelOutliningPass

gpu_kernel_outlining(composer_or_op)

gpu_launch_sink_index_computations()

Sink index computations into gpu.launch body

Argument name in MLIR CLI

gpu-launch-sink-index-computations

Pass name in TableGen

GpuLaunchSinkIndexComputationsPass

gpu_launch_sink_index_computations(composer_or_op)

gpu_map_parallel_loops()

Greedily maps loops to GPU hardware dimensions.

Argument name in MLIR CLI

gpu-map-parallel-loops

Pass name in TableGen

GpuMapParallelLoopsPass

gpu_map_parallel_loops(composer_or_op)

gpu_module_to_binary()

Transforms a GPU module into a GPU binary.

Argument name in MLIR CLI

gpu-module-to-binary

Pass name in TableGen

GpuModuleToBinaryPass

gpu_module_to_binary(composer_or_op)

inline()

Inline function calls

Argument name in MLIR CLI

inline

Pass name in TableGen

InlinerPass

inline(composer_or_op)

linalg_block_pack_matmul()

Convert linalg matmul ops to block layout and back

Argument name in MLIR CLI

linalg-block-pack-matmul

Pass name in TableGen

LinalgBlockPackMatmul

linalg_block_pack_matmul(composer_or_op)

linalg_fold_into_elementwise()

Fold transpose and broadcast ops into elementwise consumers

Argument name in MLIR CLI

linalg-fold-into-elementwise

Pass name in TableGen

LinalgFoldIntoElementwisePass

linalg_fold_into_elementwise(composer_or_op)

linalg_fold_unit_extent_dims()

Remove unit-extent dimension in Linalg ops on tensors

Argument name in MLIR CLI

linalg-fold-unit-extent-dims

Pass name in TableGen

LinalgFoldUnitExtentDimsPass

linalg_fold_unit_extent_dims(composer_or_op)

linalg_fuse_elementwise_ops()

Fuse elementwise operations on tensors

Argument name in MLIR CLI

linalg-fuse-elementwise-ops

Pass name in TableGen

LinalgElementwiseOpFusionPass

linalg_fuse_elementwise_ops(composer_or_op)

linalg_generalize_named_ops()

Convert named ops into generic ops

Argument name in MLIR CLI

linalg-generalize-named-ops

Pass name in TableGen

LinalgGeneralizeNamedOpsPass

linalg_generalize_named_ops(composer_or_op)

linalg_inline_scalar_operands()

Inline scalar operands into linalg generic ops

Argument name in MLIR CLI

linalg-inline-scalar-operands

Pass name in TableGen

LinalgInlineScalarOperandsPass

linalg_inline_scalar_operands(composer_or_op)

linalg_morph_ops()

Convert linalg ops between forms

Argument name in MLIR CLI

linalg-morph-ops

Pass name in TableGen

LinalgMorphOpsPass

linalg_morph_ops(composer_or_op)

linalg_specialize_generic_ops()

Convert generic ops back to named ops

Argument name in MLIR CLI

linalg-specialize-generic-ops

Pass name in TableGen

LinalgSpecializeGenericOpsPass

linalg_specialize_generic_ops(composer_or_op)

loop_invariant_code_motion()

Hoist loop invariant instructions outside of the loop

Argument name in MLIR CLI

loop-invariant-code-motion

Pass name in TableGen

LoopInvariantCodeMotionPass

loop_invariant_code_motion(composer_or_op)

loop_invariant_subset_hoisting()

Hoist loop invariant subset ops outside of the loop

Argument name in MLIR CLI

loop-invariant-subset-hoisting

Pass name in TableGen

LoopInvariantSubsetHoistingPass

loop_invariant_subset_hoisting(composer_or_op)

mem2reg()

Promotes memory slots into values.

Argument name in MLIR CLI

mem2reg

Pass name in TableGen

Mem2Reg

mem2reg(composer_or_op)

nvvm_attach_target()

Attaches an NVVM target attribute to a GPU Module.

Argument name in MLIR CLI

nvvm-attach-target

Pass name in TableGen

GpuNVVMAttachTarget

nvvm_attach_target(composer_or_op)

operation_type(operation_name, opts \\ [])

Creates an operation-specific Transform handle type.

param_type(type)

Creates a Transform parameter type wrapping an MLIR type.

remove_dead_values()

Remove dead values

Argument name in MLIR CLI

remove-dead-values

Pass name in TableGen

RemoveDeadValuesPass

remove_dead_values(composer_or_op)

rocdl_attach_target()

Attaches a ROCDL target attribute to a GPU Module.

Argument name in MLIR CLI

rocdl-attach-target

Pass name in TableGen

GpuROCDLAttachTarget

rocdl_attach_target(composer_or_op)

sccp()

Sparse Conditional Constant Propagation

Argument name in MLIR CLI

sccp

Pass name in TableGen

SCCPPass

sccp(composer_or_op)

simplify_depthwise_conv()

Simplify depthwise convolution.

Argument name in MLIR CLI

simplify-depthwise-conv

Pass name in TableGen

SimplifyDepthwiseConvPass

simplify_depthwise_conv(composer_or_op)

snapshot_op_locations()

Generate new locations from the current IR

Argument name in MLIR CLI

snapshot-op-locations

Pass name in TableGen

LocationSnapshot

snapshot_op_locations(composer_or_op)

spirv_attach_target()

Attaches an SPIR-V target attribute to a GPU Module.

Argument name in MLIR CLI

spirv-attach-target

Pass name in TableGen

GpuSPIRVAttachTarget

spirv_attach_target(composer_or_op)

sroa()

Scalar Replacement of Aggregates

Argument name in MLIR CLI

sroa

Pass name in TableGen

SROA

sroa(composer_or_op)

strip_debuginfo()

Strip debug info from all operations

Argument name in MLIR CLI

strip-debuginfo

Pass name in TableGen

StripDebugInfoPass

strip_debuginfo(composer_or_op)

symbol_dce()

Eliminate dead symbols

Argument name in MLIR CLI

symbol-dce

Pass name in TableGen

SymbolDCEPass

symbol_dce(composer_or_op)

symbol_privatize()

Mark symbols private

Argument name in MLIR CLI

symbol-privatize

Pass name in TableGen

SymbolPrivatizePass

symbol_privatize(composer_or_op)

topological_sort()

Sort regions without SSA dominance in topological order

Argument name in MLIR CLI

topological-sort

Pass name in TableGen

TopologicalSortPass

topological_sort(composer_or_op)

trivial_dce()

Remove trivially dead operations and blocks

Argument name in MLIR CLI

trivial-dce

Pass name in TableGen

TrivialDeadCodeEliminationPass

trivial_dce(composer_or_op)

view_op_graph()

Print Graphviz visualization of an operation

Argument name in MLIR CLI

view-op-graph

Pass name in TableGen

ViewOpGraphPass

view_op_graph(composer_or_op)

xevm_attach_target()

Attaches a XeVM target attribute to a GPU Module.

Argument name in MLIR CLI

xevm-attach-target

Pass name in TableGen

GpuXeVMAttachTarget

xevm_attach_target(composer_or_op)