Explore adding support for big endian architectures #1460
Labels
Comments
ghpvnist
added a commit
that referenced
this issue
Jul 6, 2023
Here are the following constraints:
```
(I1) operand is a tensor.
(C1) Let `E` and `E'` be the `operand` and `result` element type,
respectively and `R = rank(operand)`:
* If `num_bits(E')` = `num_bits(E)`, shape(`result`) = shape(`operand`).
* If `num_bits(E')` < `num_bits(E)`:
* `rank(result) = R+1`.
* dim(`result`, `i`) = dim(`operand`, `i`) for all `i` in [0, `R`-1].
* `dim(result, R) = num_bits(E)/num_bits(E')`.
* If `num_bits(E')` > `num_bits(E)`:
* `rank(result) = R-1`.
* dim(`result`, `i`) = dim(`operand`, `i`) for all `i` in [0, `R`-1).
* `dim(operand, R-1) = num_bits(E')/num_bits(E)`.
(C2) Conversion between complex and non-complex types is not permitted.
```
These constraints will be comprehensively covered by the following
tests:
```
I1: a) operand is not a tensor. (Covered by ODS).
C1: a) If `num_bits(E')` = `num_bits(E)`, shape(`result`) != shape(`operand`).
C1: b) If `num_bits(E')` < `num_bits(E)`: `rank(result) != R+1`.
C1: c) If `num_bits(E')` < `num_bits(E)`: dim(`result`, `i`) != dim(`operand`, `i`) for any `i` in [0, `R`-1].
C1: d) If `num_bits(E')` < `num_bits(E)`: `dim(result, R) != num_bits(E)/num_bits(E')`.
C1: e) If `num_bits(E')` > `num_bits(E)`: `rank(result) != R-1`.
C1: f) If `num_bits(E')` > `num_bits(E)`: dim(`result`, `i`) != dim(`operand`, `i`) for all `i` in [0, `R`-1).
C1: g) If `num_bits(E')` > `num_bits(E)`: `dim(operand, R-1) != num_bits(E')/num_bits(E)`.
(C2) Conversion between complex and non-complex types is permitted.
```
If we drop the "Covered by ODS" pieces, this will leave us with the
following test cases:
```
C1a: If `num_bits(E')` = `num_bits(E)`:
shape(`result`) != shape(`operand`).
C1b: If `num_bits(E')` < `num_bits(E)`:
`rank(result) != R+1`.
C1c: If `num_bits(E')` < `num_bits(E)`:
dim(`result`, `i`) != dim(`operand`, `i`) for any `i` in [0, `R`-1].
C1d: If `num_bits(E')` < `num_bits(E)`:
`dim(result, R) != num_bits(E)/num_bits(E')`.
C1e: If `num_bits(E')` > `num_bits(E)`:
`rank(result) != R-1`.
C1f: If `num_bits(E')` > `num_bits(E)`:
dim(`result`, `i`) != dim(`operand`, `i`) for all `i` in [0, `R`-1).
C1g: If `num_bits(E')` > `num_bits(E)`:
`dim(operand, R-1) != num_bits(E')/num_bits(E)`.
C2: Conversion between complex and non-complex types is permitted.
```
Notes:
* The interpreter assumes little endian representation, and we have
#1460 to address adding support for big endian architectures.
closes #1098
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This was brought up from the discussions on the OpenXLA side. Consider adding support for it in StableHLO as supported endianness for
bitcast_convertis currently implementation defined.The text was updated successfully, but these errors were encountered: