mlir.dialects._x86_transform_ops_gen¶
Attributes¶
Classes¶
Collect patterns that approximate vectorized f32 |
|
Rewrites an additive vector contraction by zeroing the accumulator input |
|
Apply patterns that rewrite bf16 and i8 matmul loops to contract packed |
|
Collect patterns to shuffle FMAs with x86 operations as operands |
|
Collect patterns to sink vector producer operations forward in a block to |
|
Collect patterns to lower a BF16 type vector.contract operation |
|
Collect patterns that tile and unroll contractions in order to prepare them |
|
Collect patterns to lower a BF16/Int8 type vector.contract operation |
|
Collect patterns to lower a F32 type vector.contract operation to a FMA. |
|
Collect patterns to lower a BF16/Int8 type vector.contract operation |
|
Functions¶
Module Contents¶
- mlir.dialects._x86_transform_ops_gen._ods_ir¶
- mlir.dialects._x86_transform_ops_gen._Buffer¶
- class mlir.dialects._x86_transform_ops_gen.ApplyMathRsqrtApproximationPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns that approximate vectorized f32
math.rsqrtoperations using the AVXvrsqrtpsinstruction (x86.avx.rsqrt) followed by one Newton-Raphson step. The inner vector dimension must be a multiple of 8.- OPERATION_NAME = 'transform.apply_patterns.x86.math_rsqrt_approximation'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyMathRsqrtApproximationPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.math_rsqrt_approximation'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_math_rsqrt_approximation(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyMathRsqrtApproximationPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyMoveAccumulatorForContractLoopPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irRewrites an additive vector contraction by zeroing the accumulator input and materializing the accumulation as a separate add after the contraction.
- OPERATION_NAME = 'transform.apply_patterns.x86.move_accumulator_for_contract_loop'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyMoveAccumulatorForContractLoopPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.move_accumulator_for_contract_loop'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_move_accumulator_for_contract_loop(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyMoveAccumulatorForContractLoopPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyPrepareVectorContractLoopsPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irApply patterns that rewrite bf16 and i8 matmul loops to contract packed VNNI operands directly, avoiding RHS unpacking. Supported unpacking uses shape casts and transposes, or split-and-join sequences. Splits may use deinterleave (binary) or transpose/extract (any width dividing the VNNI factor); joins may use interleave (binary) or slice inserts (any width dividing the VNNI factor).
- OPERATION_NAME = 'transform.apply_patterns.x86.prepare_vector_contract_loops'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyPrepareVectorContractLoopsPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.prepare_vector_contract_loops'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_prepare_vector_contract_loops(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyPrepareVectorContractLoopsPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyShuffleVectorFMAOpsPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to shuffle FMAs with x86 operations as operands such that FMAs are grouped with respect to odd/even packed index.
- OPERATION_NAME = 'transform.apply_patterns.x86.shuffle_vector_fma_ops'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyShuffleVectorFMAOpsPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.shuffle_vector_fma_ops'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_shuffle_vector_fma_ops(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyShuffleVectorFMAOpsPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplySinkVectorProducerOpsPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to sink vector producer operations forward in a block to place them immediately before their first use.
- OPERATION_NAME = 'transform.apply_patterns.x86.sink_vector_producer_ops'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplySinkVectorProducerOpsPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.sink_vector_producer_ops'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_sink_vector_producer_ops(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplySinkVectorProducerOpsPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractBF16ToFMAPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to lower a BF16 type vector.contract operation to a FMA via emulation lowering using BF16 packed operations.
- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_bf16_to_fma'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractBF16ToFMAPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_bf16_to_fma'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_vector_contract_bf16_to_fma(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyVectorContractBF16ToFMAPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractMultiLevelUnrollForAMXPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns that tile and unroll contractions in order to prepare them for the AMX nanokernel lowerings.
- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_multi_level_unroll_for_amx'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractMultiLevelUnrollForAMXPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_multi_level_unroll_for_amx'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_vector_contract_multi_level_unroll_for_amx(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyVectorContractMultiLevelUnrollForAMXPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToAMXDotProductPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to lower a BF16/Int8 type vector.contract operation to a BF16/Int8 AMX tiled dot-product.
- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_amx_dot_product'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToAMXDotProductPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_amx_dot_product'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_vector_contract_to_amx_dot_product(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyVectorContractToAMXDotProductPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToFMAPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to lower a F32 type vector.contract operation to a FMA.
- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_fma'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToFMAPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_fma'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_vector_contract_to_fma(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyVectorContractToFMAPatternsOp¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToPackedTypeDotProductPatternsOp(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None)¶
Bases:
_ods_irCollect patterns to lower a BF16/Int8 type vector.contract operation to a BF16/Int8 dot-product.
- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_packed_type_dot_product'¶
- _ODS_REGIONS = (0, True)¶
- class mlir.dialects._x86_transform_ops_gen.ApplyVectorContractToPackedTypeDotProductPatternsOpAdaptor¶
Bases:
_ods_ir- OPERATION_NAME = 'transform.apply_patterns.x86.vector_contract_to_packed_type_dot_product'¶
- mlir.dialects._x86_transform_ops_gen.apply_patterns_x86_vector_contract_to_packed_type_dot_product(*, loc: _ods_ir | None = None, ip: _ods_ir | None = None) ApplyVectorContractToPackedTypeDotProductPatternsOp¶