[Mlir] --linalg-specialize-generic-ops crashes in Casting.h:566 #122094
Assignees
Labels
Comments
|
@llvm/issue-subscribers-mlir Author: None (Emilyaxe)
git version: e4372c4
system: reproduce with: a.mlir: stack trace: |
|
@Emilyaxe: Please do not forget to add |
ok, I will add labels! |
|
@llvm/issue-subscribers-mlir-linalg Author: None (Emilyaxe)
git version: e4372c4
system: reproduce with: a.mlir: stack trace: |
|
A smaller example is (sorry if this can be made smaller): #map0 = affine_map<(i, j) -> (i, j)>
#map1 = affine_map<(i, j) -> ()>
module {
func.func @test() -> tensor<2x2xi32> {
// Two 2x2 tensor inputs.
%lhs = arith.constant dense<2> : tensor<2x2xi32>
%rhs = arith.constant dense<3> : tensor<2x2xi32>
// A scalar i32 input (the problem input).
%c0 = arith.constant 42 : i32
%init = tensor.empty() : tensor<2x2xi32>
%res =
linalg.generic
{ indexing_maps = [#map0, #map0, #map1, #map0],
iterator_types = ["parallel", "parallel"] }
// pass in scalar here
ins(%lhs, %rhs, %c0 : tensor<2x2xi32>, tensor<2x2xi32>, i32)
outs(%init : tensor<2x2xi32>)
{
^bb0(%x: i32, %y: i32, %scalar: i32, %out: i32):
%add = arith.addi %x, %y : i32
%sum = arith.addi %add, %scalar : i32
linalg.yield %sum : i32
} -> tensor<2x2xi32>
// Return the computed tensor.
return %res : tensor<2x2xi32>
}
}The issue here being that #map0 = affine_map<(i, j) -> (i, j)>
#map1 = affine_map<(i, j) -> ()>
module {
func.func @test() -> tensor<2x2xi32> {
// Two 2x2 tensor inputs.
%lhs = arith.constant dense<2> : tensor<2x2xi32>
%rhs = arith.constant dense<3> : tensor<2x2xi32>
// A scalar i32 input.
%c0 = arith.constant 42 : i32
// create tensor
%ct0 = tensor.from_elements %c0 : tensor<i32>
// ...
ins(%lhs, %rhs, %ct0 : tensor<2x2xi32>, tensor<2x2xi32>, tensor<i32>)
// ...
}
}I was having a look through the code and this validates the operands: // LinalgInterfaces.cpp
for (OpOperand &opOperand : linalgOp->getOpOperands()) {
AffineMap indexingMap = linalgOp.getMatchingIndexingMap(&opOperand);
// Symbols disallowed.
if (indexingMap.getNumSymbols() != 0)
return op->emitOpError("unexpected symbols in indexing_map #")
<< opOperand.getOperandNumber();
// Domain must be consistent.
unsigned numLoops = linalgOp.getNumLoops();
if (indexingMap.getNumDims() != numLoops)
return op->emitOpError("expected indexing_map #")
<< opOperand.getOperandNumber() << " to have " << numLoops
<< " dim(s) to match the number of loops";
int64_t rank = linalgOp.getRank(&opOperand);
if (indexingMap.getNumResults() != rank)
return op->emitOpError("expected operand rank (")
<< rank << ") to match the result rank of indexing_map #"
<< opOperand.getOperandNumber() << " ("
<< indexingMap.getNumResults() << ")";
}The issue is here the // DecomposeGenericByUnfoldingPermutation.cpp
if (llvm::any_of(op->getOpOperands(), [](OpOperand &oper) {
auto opType = cast<RankedTensorType>(oper.get().getType());
return ShapedType::isDynamicShape(opType.getShape());
}))
return failure();It attempts to cast I am new to MLIR so not sure what the best cause of action would be, but where the checks are done in |
|
I think we should add extra check for RankedTensorType operand. CC @javedabsar1 |
|
@CoTinker could you assign this to me and I will add the check? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
git version: e4372c4
system:
Ubuntu 18.04.6 LTSreproduce with:
mlir-opt a.mlir --linalg-specialize-generic-opsa.mlir:
stack trace:
The text was updated successfully, but these errors were encountered: