Segfault on convert_simplify optimization with complex numbers
#695
Comments
|
Seems like julia> Reactant.Compiler.run_pass_pipeline_on_source("""
module {
func.func @main(%arg0: tensor<2xcomplex<f64>>) -> (tensor<complex<f64>>, tensor<2xcomplex<f64>>) {
%0 = stablehlo.transpose %arg0, dims = [0] : (tensor<2xcomplex<f64>>) -> tensor<2xcomplex<f64>>
%c = stablehlo.constant dense<1> : tensor<2xi64>
%1 = stablehlo.convert %c : (tensor<2xi64>) -> tensor<2xcomplex<f64>>
%2 = stablehlo.dot_general %0, %1, contracting_dims = [0] x [0] : (tensor<2xcomplex<f64>>, tensor<2xcomplex<f64>>) -> tensor<complex<f64>>
%3 = stablehlo.transpose %0, dims = [0] : (tensor<2xcomplex<f64>>) -> tensor<2xcomplex<f64>>
return %2, %3 : tensor<complex<f64>>, tensor<2xcomplex<f64>>
}
}""", "enzyme-hlo-generate-td{patterns=convert_simplify},transform-interpreter")but with julia> Reactant.Compiler.run_pass_pipeline_on_source("""
module {
func.func @main(%arg0: tensor<2xf64>) -> (tensor<f64>, tensor<2xf64>) {
%0 = stablehlo.transpose %arg0, dims = [0] : (tensor<2xf64>) -> tensor<2xf64>
%c = stablehlo.constant dense<1> : tensor<2xi64>
%1 = stablehlo.convert %c : (tensor<2xi64>) -> tensor<2xf64>
%2 = stablehlo.dot_general %0, %1, contracting_dims = [0] x [0] : (tensor<2xf64>, tensor<2xf64>) -> tensor<f64>
%3 = stablehlo.transpose %0, dims = [0] : (tensor<2xf64>) -> tensor<2xf64>
return %2, %3 : tensor<f64>, tensor<2xf64>
}
}""", "enzyme-hlo-generate-td{patterns=convert_simplify},transform-interpreter")
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op) {
%0 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op
transform.apply_patterns to %0 {
transform.apply_patterns.enzyme_hlo.convert_simplify
} : !transform.any_op
transform.yield
}
func.func @main(%arg0: tensor<2xf64>) -> (tensor<f64>, tensor<2xf64>) {
%cst = stablehlo.constant dense<1.000000e+00> : tensor<2xf64>
%0 = stablehlo.transpose %arg0, dims = [0] : (tensor<2xf64>) -> tensor<2xf64>
%1 = stablehlo.dot_general %0, %cst, contracting_dims = [0] x [0] : (tensor<2xf64>, tensor<2xf64>) -> tensor<f64>
%2 = stablehlo.transpose %0, dims = [0] : (tensor<2xf64>) -> tensor<2xf64>
return %1, %2 : tensor<f64>, tensor<2xf64>
}
} |
mofeing
changed the title
convert_simplify optimization with complex numbers
|
The smaller reproducer provided by Sergio hits an assertion in a debug build of Reactant: The assertion is https://github.com/EnzymeAD/Enzyme-JAX/blob/d694d09a33814a83abaef9d1ef924614ab59e4c5/src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp#L1343 if (auto floatType = dyn_cast<FloatType>(elemType))
bitWidth = floatType.getWidth();
assert(bitWidth != -1 && "expect integer or float"); |
|
ironically x/ref openxla/stablehlo#2709 cc @GleasonK |
Merged
Merged
Merged
Merged
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
CC @mofeing
We found the following unexpected behavior when performing
dot_generaloperation over two vectors of different types.The text was updated successfully, but these errors were encountered: