mlir::VectorTransferToSCFOptions Struct Reference

When lowering an N-d vector transfer op to an (N-1)-d vector transfer op, a temporary buffer is created through which individual (N-1)-d vector are staged. More...

#include "mlir/Conversion/VectorToSCF/VectorToSCF.h"

Public Member Functions
VectorTransferToSCFOptions &	setTargetRank (unsigned r)
VectorTransferToSCFOptions &	enableLowerTensors (bool l=true)
VectorTransferToSCFOptions &	enableFullUnroll (bool u=true)

Public Attributes
unsigned	targetRank = 1
	Minimal rank to which vector transfer are lowered. More...
bool	lowerTensors = false
	Allows vector transfers that operated on tensors to be lowered (this is an uncommon alternative). More...
bool	unroll = false
	Triggers full unrolling (vs iterating with a loop) during transfer to scf. More...

Detailed Description

When lowering an N-d vector transfer op to an (N-1)-d vector transfer op, a temporary buffer is created through which individual (N-1)-d vector are staged.

This pattern can be applied multiple time, until the transfer op is 1-d. This is consistent with the lack of an LLVM instruction to dynamically index into an aggregate (see the Vector dialect lowering to LLVM deep dive).

An instruction such as:

vector.transfer_write %vec, %A[%a, %b, %c] :
  vector<9x17x15xf32>, memref<?x?x?xf32>

Lowers to pseudo-IR resembling (unpacking one dimension):

%0 = alloca() : memref<vector<9x17x15xf32>>
store %vec, %0[] : memref<vector<9x17x15xf32>>
%1 = vector.type_cast %0 :
  memref<vector<9x17x15xf32>> to memref<9xvector<17x15xf32>>
affine.for %I = 0 to 9 {
  %dim = dim %A, 0 : memref<?x?x?xf32>
  %add = affine.apply %I + %a
  %cmp = arith.cmpi "slt", %add, %dim : index
  scf.if %cmp {
    %vec_2d = load %1[%I] : memref<9xvector<17x15xf32>>
    vector.transfer_write %vec_2d, %A[%add, %b, %c] :
      vector<17x15xf32>, memref<?x?x?xf32>

When applying the pattern a second time, the existing alloca() operation is reused and only a second vector.type_cast is added.

Definition at line 52 of file VectorToSCF.h.

Member Function Documentation

enableFullUnroll()

VectorTransferToSCFOptions& mlir::VectorTransferToSCFOptions::enableFullUnroll ( bool  u = true )

inline

Definition at line 68 of file VectorToSCF.h.

References unroll.

enableLowerTensors()

VectorTransferToSCFOptions& mlir::VectorTransferToSCFOptions::enableLowerTensors ( bool  l = true )

inline

Definition at line 62 of file VectorToSCF.h.

References lowerTensors.

setTargetRank()

VectorTransferToSCFOptions& mlir::VectorTransferToSCFOptions::setTargetRank ( unsigned  r )

inline

Definition at line 55 of file VectorToSCF.h.

References targetRank.

Member Data Documentation

lowerTensors

bool mlir::VectorTransferToSCFOptions::lowerTensors = false

Allows vector transfers that operated on tensors to be lowered (this is an uncommon alternative).

Definition at line 61 of file VectorToSCF.h.

Referenced by enableLowerTensors().

targetRank

unsigned mlir::VectorTransferToSCFOptions::targetRank = 1

Minimal rank to which vector transfer are lowered.

Definition at line 54 of file VectorToSCF.h.

Referenced by setTargetRank().

unroll

bool mlir::VectorTransferToSCFOptions::unroll = false

Triggers full unrolling (vs iterating with a loop) during transfer to scf.

Definition at line 67 of file VectorToSCF.h.

Referenced by enableFullUnroll().

---

The documentation for this struct was generated from the following file:

• include/mlir/Conversion/VectorToSCF/VectorToSCF.h