Add spec for UniformQuantizeOp #531
Assignees
Labels
Comments
|
Will be handled as part of the quantization speccing effort (#588). |
sdasgup3
added a commit
that referenced
this issue
Jun 19, 2023
fixes #531 fixes #530 ## Summary The PR proposes the specification for `uniform.quantize` and `uniform.dequantize` ops. The specification of `uniform.quantize` also captures the re-quantization conversions from quantized tensor to quantized tensors. Please let me know your feedback. ### Working notes on following the `reference_checklist.md` and `spec_checkilist.md` #### uniform_dequantize We have the following constraints from the spec ``` (I1) `operand` is a quantized tensor (C1) `shape(operand) = shape(result)`. (C2) `element_type(result) = expressed_type(operand)`. ``` These constraints will be comprehensively covered by the following tests: ``` I1: a) `operand: quantized tensors`. (Covered by ODS). C1: a) `shape(operand) != shape(result)`. (Covered by ODS) C2: a) `element_type(result) != expressed_type(operand)`. ``` If we drop the "Covered by ODS" pieces, this will leave us with the following test cases: ``` C2: a) `element_type(result) != expressed_type(operand)`. ``` We already has a type inference test to cover the above. #### uniform_quantize We have the following constraints from the spec ``` (I1) `operand: tensor of floating-point or quantized type`. (C1) `shape(operand) = shape(result)`. (C2) `expressed_type(result) = is_float(operand) ? element_type(operand) : expressed_type(operand)`. ``` These constraints will be comprehensively covered by the following tests: ``` I1: a) `operand: quantized tensors`. (Covered by ODS). C1: a) `shape(operand) != shape(result)`. (Covered by ODS) C2: a) if_float(operand): `expressed_type(result) != element_type(operand)`. b) if_quantized(operand): `expressed_type(result) != expressed_type(operand)`. ``` If we drop the "Covered by ODS" pieces, this will leave us with the following test cases: ``` C2: a) if_float(operand): `expressed_type(result) != element_type(operand)`. b) if_quantized(operand): `expressed_type(result) != expressed_type(operand)`. ``` The above will be covered as part of #1603.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Request description
To spec
stablehlo.uniform_quantizeThe text was updated successfully, but these errors were encountered: