- refactor to take a function to generate the propore BaseValue to store in
its output
- refactor BaseTracer to force inheriting tracers to indicate how to build
a ConstantInput tracer
- remove "as import" for intermediate in hnumpy/tracing.py
- update compile to manage python dtypes
- rename make_integer_to_hold_ints to make_integer_to_hold
- accept any values as input as we don't know which type this function will
be called with
- rename get_bits_to_represent_int to
get_bits_to_represent_value_as_integer
- add _is_encrypted to BaseValue
- remove EncryptedValue and ClearValue classes
- add a ScalarValue class
- add two helpers EncryptedValue and ClearValue which create a ScalarValue
either encrypted or not when passed a data_type
- rename to mix_scalar_values_determine_holding_dtype
- change typing
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
* feat: let the dev give a useful name for ArbitraryFunction
might be useful to debug or understand what happens
closes#144
Co-authored-by: Benoit Chevallier-Mames <benoitchevalliermames@zama.ai>
* feat(mlir): conversion from HDKIR to MLIR
* feat(mlir): support ir.Sub and ir.Mul
- better type conversion from HDK to MLIR
- Context management inside the converter class
- better handling of input type in conversion functions
* refactor(mlir): use input and output from OPGraph
Co-authored-by: Arthur Meyre <arthur.meyre@zama.ai>
* feat(mlir): eint-int subtractions
* feat(mlir): adhere to spec for supported ops
* feat(OPGraph): getters for ordered inputs/outputs
+ formatting
* tests(mlir): test converion via compiler roundtrip
* fix(mlir): flip operands on int_eint sym ops
* feat(mlir): check that the outputs are unsigned
* feat(mlir): set bit_width of all nodes to the max
This is currently required as the compiler is already assuming this.
Could be removed from HDK when the compiler can do it on its own
* feat: value_is_integer + CRs
disable some linting errors
* tests: update compile tests + coverage
* refactor: reorganize mlir package + better doc
* doc: conformance with pydocstyle
Co-authored-by: Arthur Meyre <arthur.meyre@zama.ai>
- start tracing numpy.rint and manage dtypes
- update BaseTracer to accept iterables as inputs, because NPTracer does
not get list givent the way numpy sends arguments to the functions
- add function to convert a type from the project to a numpy dtype
- add function to manage numpy ufunc output dtypes
- add a check for integers to have positive bit_width
- add a check for floats to only accept 32 and 64 bits
- this allows to make sure each node has a proper implementation
- move op_args and op_kwargs in ArbitraryFunction only
- update BaseTracer accordingly
- update ConstantInput to manage floats
- update OPGraph update_values_with_bounds to manage floats
- update test code to manage cases where output could be a float
- add test cases with float inputs
- this allows to have tighter data types by sticking to the smallest types
able to represent the ranges passed as argument
- update test_dataset_eval to check the output Value's data_type is updated
- this input index will be useful for MLIR/lower level conversions
- it represents the input index of an Input node when considering the
traced function signature
- update code preparing function parameters to keep signature order
- sometimes arguments are re-ordered in the id written by pytest to sdtout
which makes identifying the failing case hard, this solves the issue for
one test function
- allows to have a generic way of propagating types instead of deepcopying
input values and types in IR nodes
- supports only Integers for now
- opset checks will not be performed in those functions to keep knowledge
required on the opset to the MLIR conversion step or extra check steps
- use the mix_values_determine_holding_dtype in intermediate nodes where
appropriate
tests: add the possibility to test more binary operations
avoid copy-paste for binary ops, with just an option if the operation is commutative or not
refs #41, #42
