| Issue |
174550
|
| Summary |
[MLIR] `-linalg-generalize-named-ops -linalg-specialize-generic-ops` remove `cast_unsigned` in linalg.matmul, leading to Inconsistent Results
|
| Labels |
mlir
|
| Assignees |
|
| Reporter |
BealZephyr
|
**test commit**: [5cfd02](https://github.com/llvm/llvm-project/commit/5cfd02f44a43a2e2a085a633b022a62f64ba2b93)
## Description:
With `cast_unsigned`, the values are first converted to unsigned integers:
The tensor values `-2` in `i16` are interpreted as `65534` in `uint16`.
The matrix multiplication is performed using these unsigned values:
`65534 * 3 = 196602`
`196602 * 2 = 393204`
So, the resulting matrix will be filled with the value `393204`.
Adding` -linalg-generalize-named-ops and -linalg-specialize-generic-ops`, the effect of cast_unsigned seems to be bypassed, and the operation uses the original signed values, leading to the result -12
## test case
```
module {
func.func private @printMemrefI32(tensor<*xi32>)
func.func @main() {
%0 = "tosa.const"() <{values = dense<-2> : tensor<4x2xi16>}> : () -> tensor<4x2xi16>
%1 = "tosa.const"() <{values = dense<3> : tensor<2x8xi64>}> : () -> tensor<2x8xi64>
%2 = "tosa.const"() <{values = dense<0> : tensor<4x8xi32>}> : () -> tensor<4x8xi32>
%3 = linalg.matmul {cast = #linalg.type_fn<cast_unsigned>} ins(%0, %1 : tensor<4x2xi16>, tensor<2x8xi64>) outs(%2 : tensor<4x8xi32>) -> tensor<4x8xi32>
%cast = tensor.cast %3 : tensor<4x8xi32> to tensor<*xi32>
call @printMemrefI32(%cast) : (tensor<*xi32>) -> ()
return
}
}
```
## Command:
### 1. without `-linalg-generalize-named-ops -linalg-specialize-generic-ops`
#### cmd
```
mlir-opt test.mlir -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries=1" -convert-scf-to-cf -convert-linalg-to-loops \
-convert-scf-to-cf -convert-to-llvm | mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project-latest/build/lib/libmlir_runner_utils.so
```
#### output:
```
[[393204, 393204, 393204, 393204, 393204, 393204, 393204, 393204],
[393204, 393204, 393204, 393204, 393204, 393204, 393204, 393204],
[393204, 393204, 393204, 393204, 393204, 393204, 393204, 393204],
[393204, 393204, 393204, 393204, 393204, 393204, 393204, 393204]]
```
### 2. With `-linalg-generalize-named-ops -linalg-specialize-generic-ops`
#### cmd
```
mlir-opt test.mlir -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries=1" -convert-scf-to-cf -linalg-generalize-named-ops \
-linalg-specialize-generic-ops -convert-linalg-to-loops -convert-scf-to-cf -convert-to-llvm | mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project-latest/build/lib/libmlir_runner_utils.so
```
#### output:
```
[[-12, -12, -12, -12, -12, -12, -12, -12],
[-12, -12, -12, -12, -12, -12, -12, -12],
[-12, -12, -12, -12, -12, -12, -12, -12],
[-12, -12, -12, -12, -12, -12, -12, -12]
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
