All results in code compiled by zamacompiler are passed as return
values, which means that all tensors passed as function arguments are
constant inputs that are never written.
This patch changes the arguments used as data pointers for input
tensors in `JITLambda::Arguments::setArg()` from `void*` to `const
void*` to emphasize their use as inputs and to allow for constant
arrays to be passed as function inputs.
By default, `mlir::SourceMgrDiagnosticVerifierHandler` used by
`CompilerEngine::compile()` prints parse errors to
`llvm::errs()`. This makes it impossible for a caller of
`CompilerEngine::compile()` to process parse errors or to suppress the
emission of error messages to the standard error stream altogether.
This change captures parse errors in a string via a string-backed
output stream and forwards the error message in the `llvm::Error`
instance of the return value.
The code in `lib/CAPI/Support/CompilerEngine.cpp` invokes several
functions returning an `llvm::Expected<T>`. When those fail, the error
message retrieved from the error object the `llvm::Expected<T>`
instance is written to the standard error stream via
`mlir::zamalang::log_error()` and an exception with a more generic
error message is thrown.
This causes errors to show up on the standard error stream in tests
generating errors on purpose and checking them, e.g.:
```
tests/python/test_compiler_engine.py::test_compile_invalid[not
@main] Compilation failed: cannot find the function for generate
client parameters PASSED
```
This patch forwards the error message from an `llvm::Expected<T>`
instance in a runtime exception rather than writing it to the standard
error stream. Since exceptions are properly caught by the tests, no
errors show up during testing.
This commit contains several incremental improvements towards a clear
interface for lambdas:
- Unification of static and JIT compilation by using the static
compilation path of `CompilerEngine` within a new subclass
`JitCompilerEngine`.
- Clear ownership for compilation artefacts through
`CompilationContext`, making it impossible to destroy objects used
directly or indirectly before destruction of their users.
- Clear interface for lambdas generated by the compiler through
`JitCompilerEngine::Lambda` with a templated call operator,
encapsulating otherwise manual orchestration of `CompilerEngine`,
`JITLambda`, and `CompilerEngine::Argument`.
- Improved error handling through `llvm::Expected<T>` and proper
error checking following the conventions for `llvm::Expected<T>`
and `llvm::Error`.
Co-authored-by: youben11 <ayoub.benaissa@zama.ai>
Composing error messages for `llvm::Error` is either done by using
`llvm::createStringError()` with an appropriate format string and
arguments or by writing to a `std::string`-backed
`llvm::raw_string_ostream` and passing the result to
`llvm::make_error<llvm::StringError>()` verbatim.
The new class `StreamStringError` encapsulates the latter solution
into a class with an appropriate stream operator and implicit cast
operators to `llvm::Error` and `llvm::Expected`.
Example usage:
llvm::Error foo(int i, size_t s, ...) {
...
if(...) {
return StreamStringError()
<< "Some error message with an integer: "
<< i << " and a size_t: " << s;
}
...
}
When determining the maximum MANP and precision, the `MaxMANPPass`
only takes into account results generated by an operation, but ignores
function parameters. However, encrypted function parameters are
assumed to have a MANP value of 1 and can have an arbitrary
precision. This patch takes into account function arguments by using
their default MANP values and the extracting their precision.
Add support for `tensor.extract` operations in the MANP pass. This
currently only supports extract operations on tensors of encrypted
integers, which are passed as function arguments, e.g.:
func @extract_ith(%t: tensor<10x!HLFHE.eint<5>>, %i: index) -> !HLFHE.eint<5>{
%c = tensor.extract %t[%i] : tensor<10x!HLFHE.eint<5>>
return %c : !HLFHE.eint<5>
}
- Docker image to build wheels for linux_x86_64 CPython 3.[8,9,10] with
GLIBC >= 2.24
- Specify which Python to use in Makefile
- Fix cmake build to handle when libpython isn't available (cmake>3.18)
We decide to make this choice as they are issue to crate tensor of custom integer type in python.
+ we don't do the integer extension before convert to the concrete CAPI that requires i64
LLVM errors should be handled/consumed. Creating a new one and leaving
the previous one alive will crash the compiler. Whenever we don't want a
crash (e.g. logging the error is enough), but still wanna continue the
execution, we can just consume it.
This replaces the default FHE circuit constrains (maximum encrypted
integer width of 7 bits and a Minimal Arithmetic Noise Padding of 10
with the results of the `MaxMANP` pass, which determines these values
automatically from the input program.
Since the maximum encrypted integer width and the maximum value for
the Minimal Arithmetic Noise Padding can only be derived from HLFHE
operations, the circuit constraints are determined automatically by
`zamacompiler` only if the option `--entry-dialect=hlfhe` was
specified.
For lower-level dialects, `zamacompiler` has been provided with the
options `--assume-max-eint-precision=...` and `--assume-max-manp=...`
that allow a user to specify the values for the maximum required
precision and maximum values for the Minimal Arithmetic Noise Padding.
A value of 1 is perfectly valid for the 2-norm of the constant vector
of a dot operation. Hence, the log2 of that value of 0 is also
perfectly valid. However, the offset of -1 in `getV0Parameter` applied
to that value yields an offset into a static lookup table of -1, which
triggers a segmentation fault.
This patch removes the offset of -1 for the log2 value of the 2-norm
in `getV0Parameter`.
This pass calculates the squared Minimal Arithmetic Noise Padding
(MANP) for each operation using the MANP pass and extracts the maximum
(non-squared) Minimal Arithmetic Noise Padding and the maximum
ecrypted integer width from.
