This adds support for the tiling of `linalg.generic` operations that
have only parallel iterators or only parallel iterators and a single
reduction dimension via the linalg tiling infrastructure (i.e.,
`mlir::linalg::tileToForallOpUsingTileSizes()` and
`mlir::linalg::tileReductionUsingForall()`).
This allows for the tiling of FHELinalg operations by first replacing
them with appropriate `linalg.generic` oeprations and then invoking
the tiling pass in the pipeline. In order for the tiling to take
place, tile sizes must be specified using the `tile-sizes` operation
attribute, either directly for `linalg.generic` operations or
indirectly for the FHELinalg operation, e.g.,
"FHELinalg.matmul_eint_int"(%a, %b) { "tile-sizes" = [0, 0, 7] } : ...
Tiling of operations with a reduction dimension is currently limited
to tiling of the reduction dimension, i.e., the tile sizes for the
parallel dimensions must be zero.
Some of the tests use lookup tables whose numbers of elements do not
match the sizes of the polynoms of the bootstrap operations they are
passed to. This commit replaces these lookup tables with tables of the
right size.
This commit:
+ Adds support for a protocol which enables inter-op between concrete,
tfhe-rs and potentially other contributors to the fhe ecosystem.
+ Gets rid of hand-made serialization in the compiler, and
client/server libs.
+ Refactors client/server libs to allow more pre/post processing of
circuit inputs/outputs.
The protocol is supported by a definition in the shape of a capnp file,
which defines different types of objects among which:
+ ProgramInfo object, which is a precise description of a set of fhe
circuit coming from the same compilation (understand function type
information), and the associated key set.
+ *Key objects, which represent secret/public keys used to
encrypt/execute fhe circuits.
+ Value object, which represent values that can be transferred between
client and server to support calls to fhe circuits.
The hand-rolled serialization that was previously used is completely
dropped in favor of capnp in the whole codebase.
The client/server libs, are refactored to introduce a modular design for
pre-post processing. Reading the ProgramInfo file associated with a
compilation, the client and server libs assemble a pipeline of
transformers (functions) for pre and post processing of values coming in
and out of a circuit. This design properly decouples various aspects of
the processing, and allows these capabilities to be safely extended.
In practice this commit includes the following:
+ Defines the specification in a concreteprotocol package
+ Integrate the compilation of this package as a compiler dependency
via cmake
+ Modify the compiler to use the Encodings objects defined in the
protocol
+ Modify the compiler to emit ProgramInfo files as compilation
artifact, and gets rid of the bloated ClientParameters.
+ Introduces a new Common library containing the functionalities shared
between the compiler and the client/server libs.
+ Introduces a functional pre-post processing pipeline to this common
library
+ Modify the client/server libs to support loading ProgramInfo objects,
and calling circuits using Value messages.
+ Drops support of JIT.
+ Drops support of C-api.
+ Drops support of Rust bindings.
Co-authored-by: Nikita Frolov <nf@mkmks.org>
Instead of having one `getSQManp` implementation per op with a lot of repetition, the noise
calculation is now modular.
- Ops that implements`UnaryEint`/`BinaryInt`/`BinaryEint` interfaces share the operand noise
presence check.
- For many scalar ops no further calculation is needed. If it's not the case, an op can override
`sqMANP`.
- Integer operand types lookups are abstracted into `BinaryInt::operandIntType()`
- Finding largest operand value for a type is abstracted into `BinaryInt::operandMaxConstant`
- Noise calculation for matmul ops is simplified and it's now general enough to work for
`matmul_eint_int`, `matmul_int_eint` and `dot_eint_int` at once.
The current batching pass only supports batching of operations that
have a single batchable operand, that can only be batched in one way
and that operate on scalar values. However, this does not allow for
efficient batching of all arithmetic operations in TFHE, since these
are often applied to pairs of scalar values from tensors, to tensors
and scalars or to tensors that can be grouped in higher-order tensors.
This commit introduces three new features for batching:
1. Support of multiple batchable operands
The operation interface for batching now allows for the
specification of multiple batchable operands. This set can be
composed of any subset of an operation's operands, i.e., it is
not limited to sets of operands with contiguous operand indexes.
2. Support for multiple batching variants
To account for multiple kinds of batching, the batching operation
interface `BatchableOpInterface` now supports variants. The
batching pass attempts to batch an operation by trying the
batching variants expressed via the interface in order until it
succeeds.
3. Support for batching of tensor values
Some operations that could be batched already operate on tensor
values. The new batching pass detects those patterns and groups
the batchable tensors' values into higher-dimensional tensors.
This commit brings support for multiple secret keys in the TFHE
dialect. In particular, a parameterized `TFHE` circuit can now be
given as input, with any combination of (semantically valid) of
ks/bs/woppbs mixing different secret keys, and compiled down to a
valid executable function, with server keys properly looked up.
Secret keys are now stateful objects which can be:
-> none/unparameterized (syntax `sk?`): The keys are in state after
the lowering from the `FHE` dialect.
-> parameterized (syntax `sk<identifier, polysize, dimension>`): The
keys were parameterized, either by user or by the optimizer. The
`identifier` field can be used to disambiguate two keys with same
`polysize` and `dimension`.
-> normalized (syntax `sk[index]<polysize, dimension>`): The keys were
attached to their index in the list of keys in the runtime context.
The _normalization_ of key indices also acts on the ksk, bsk and pksk,
which are given indices in the same spirit now.
Finally, in order to allow parameterized `TFHE` circuit to be given
as input and compiled down to executable functions, we added a way to
pass the encodings that are used to encode/decode the circuit
inputs/outputs. In the case of a compilation from the `FHE` dialect,
those informations are automatically extracted from the higher level
informations available in this dialect.
Re-introduce the previously deleted batching tests for BConcrete as
tests for TFHE with the addition of a new test, checking that
non-batchable operands generated by pure operations are hoisted.
This commit rebases the compiler onto commit 465ee9bfb26d from
llvm-project with locally maintained patches on top, i.e.:
* 5d8669d669ee: Fix the element alignment (size) for memrefCopy
* 4239163ea337: fix: Do not fold the memref.subview if the offset are
!= 0 and strides != 1
* 72c5decfcc21: remove github stuff from llvm
* 8d0ce8f9eca1: Support arbitrary element types in named operations
via attributes
* 94f64805c38c: Copy attributes of scf.for on bufferization and make
it an allocation hoisting barrier
Main upstream changes from llvm-project that required modification of
concretecompiler:
* Switch to C++17
* Various changes in the interfaces for linalg named operations
* Transition from `llvm::Optional` to `std::optional`
* Use of enums instead of string values for iterator types in linalg
* Changed default naming convention of getter methods in
ODS-generated operation classes from `some_value()` to
`getSomeValue()`
* Renaming of Arithmetic dialect to Arith
* Refactoring of side effect interfaces (i.e., renaming from
`NoSideEffect` to `Pure`)
* Re-design of the data flow analysis framework
* Refactoring of build targets for Python bindings
* Refactoring of array attributes with integer values
* Renaming of `linalg.init_tensor` to `tensor.empty`
* Emission of `linalg.map` operations in bufferization of the Tensor
dialect requiring another linalg conversion pass and registration
of the bufferization op interfaces for linalg operations
* Refactoring of the one-shot bufferizer
* Necessity to run the expand-strided-metadata, affine-to-std and
finalize-memref-to-llvm passes before converson to the LLVM
dialect
* Renaming of `BlockAndValueMapping` to `IRMapping`
* Changes in the build function of `LLVM::CallOp`
* Refactoring of the construction of `llvm::ArrayRef` and
`llvm::MutableArrayRef` (direct invocation of constructor instead
of builder functions for some cases)
* New naming conventions for generated SSA values requiring rewrite
of some check tests
* Refactoring of `mlir::LLVM::lookupOrCreateMallocFn()`
* Interface changes in generated type parsers
* New dependencies for to mlir_float16_utils and
MLIRSparseTensorRuntime for the runtime
* Overhaul of MLIR-c deleting `mlir-c/Registration.h`
* Deletion of library MLIRLinalgToSPIRV
* Deletion of library MLIRLinalgAnalysis
* Deletion of library MLIRMemRefUtils
* Deletion of library MLIRQuantTransforms
* Deletion of library MLIRVectorToROCDL
