| Issue |
87050
|
| Summary |
[MLIR][Arith] Convert `arith.truncf` with rounding mode to SPIR-V
|
| Labels |
enhancement,
mlir:spirv,
mlir
|
| Assignees |
|
| Reporter |
victor-eds
|
#86152 added an optional rounding mode attribute to `arith.truncf`. Due to missing infrastructure in SPIR-V, conversion to the SPIR-V dialect was not implemented as part of that PR.
SPIR-V accepts specifying an operation rounding mode using the `FPRoundingMode` decoration, receiving an [FP rounding mode](https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#FP_Rounding_Mode) value. The decoration key is present in the dialect already, but we're missing a representation for the FP rounding mode values. Once we have that, we could convert:
```mlir
func.func @experimental_constrained_fptrunc(%arg0 : f64) {
%0 = arith.truncf %arg0 tonearesteven : f64 to f32
%1 = arith.truncf %arg0 downward : f64 to f32
%2 = arith.truncf %arg0 upward : f64 to f32
%3 = arith.truncf %arg0 towardzero : f64 to f32
return
}
```
into:
```mlir
func.func @experimental_constrained_fptrunc(%arg0 : f64) {
%0 = spirv.FConvertOp %arg0 {FPRoundingMode = #spirv.FPRoundingMode<RTE>} : f64 to f32
%1 = spirv.FConvertOp %arg0 {FPRoundingMode = #spirv.FPRoundingMode<RTN>} : f64 to f32
%2 = spirv.FConvertOp %arg0 {FPRoundingMode = #spirv.FPRoundingMode<RTP>} : f64 to f32
%3 = spirv.FConvertOp %arg0 {FPRoundingMode = #spirv.FPRoundingMode<RTZ>} : f64 to f32
return
}
```
Note there is no representation of `tonearestaway` in SPIR-V, so conversion should fail in that case.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
