mirror of
https://github.com/zama-ai/concrete.git
synced 2026-02-08 19:44:57 -05:00
d41d14dbb8
this is a first commit to support operations on U64 by decomposing them into smaller chunks (32 chunks of 2 bits). This commit introduce the lowering pass that will be later populated to support other operations.
223 lines
9.1 KiB
C++
223 lines
9.1 KiB
C++
// Part of the Concrete Compiler Project, under the BSD3 License with Zama
|
|
// Exceptions. See
|
|
// https://github.com/zama-ai/concrete-compiler-internal/blob/main/LICENSE.txt
|
|
// for license information.
|
|
|
|
#include <mlir/Dialect/Affine/IR/AffineOps.h>
|
|
#include <mlir/Dialect/Arithmetic/IR/Arithmetic.h>
|
|
#include <mlir/Dialect/Func/IR/FuncOps.h>
|
|
#include <mlir/Dialect/Tensor/IR/Tensor.h>
|
|
#include <mlir/IR/PatternMatch.h>
|
|
#include <mlir/Transforms/DialectConversion.h>
|
|
|
|
#include <concretelang/Conversion/Utils/GenericOpTypeConversionPattern.h>
|
|
#include <concretelang/Conversion/Utils/Legality.h>
|
|
#include <concretelang/Conversion/Utils/ReinstantiatingOpTypeConversion.h>
|
|
#include <concretelang/Dialect/FHE/IR/FHEOps.h>
|
|
#include <concretelang/Dialect/FHE/IR/FHETypes.h>
|
|
#include <concretelang/Dialect/FHE/Transforms/BigInt/BigInt.h>
|
|
#include <concretelang/Support/Constants.h>
|
|
|
|
namespace mlir {
|
|
namespace concretelang {
|
|
|
|
/// Construct a table lookup to extract the carry bit
|
|
mlir::Value getTruthTableCarryExtract(mlir::PatternRewriter &rewriter,
|
|
mlir::Location loc,
|
|
unsigned int chunkSize,
|
|
unsigned int chunkWidth) {
|
|
auto tableSize = 1 << chunkSize;
|
|
std::vector<llvm::APInt> values;
|
|
values.reserve(tableSize);
|
|
for (auto i = 0; i < tableSize; i++) {
|
|
if (i < 1 << chunkWidth)
|
|
values.push_back(llvm::APInt(1, 0, false));
|
|
else
|
|
values.push_back(llvm::APInt(1, 1, false));
|
|
}
|
|
auto truthTableAttr = mlir::DenseElementsAttr::get(
|
|
mlir::RankedTensorType::get({tableSize}, rewriter.getIntegerType(64)),
|
|
values);
|
|
auto truthTable =
|
|
rewriter.create<mlir::arith::ConstantOp>(loc, truthTableAttr);
|
|
return truthTable.getResult();
|
|
}
|
|
|
|
namespace {
|
|
|
|
namespace typing {
|
|
|
|
/// Converts `FHE::ChunkedEncryptedInteger` into a tensor of
|
|
/// `FHE::EncryptedInteger`.
|
|
mlir::RankedTensorType
|
|
convertChunkedEint(mlir::MLIRContext *context,
|
|
FHE::ChunkedEncryptedIntegerType chunkedEint,
|
|
unsigned int chunkSize, unsigned int chunkWidth) {
|
|
auto eint = FHE::EncryptedIntegerType::get(context, chunkSize);
|
|
auto bigIntWidth = chunkedEint.getWidth();
|
|
assert(bigIntWidth % chunkWidth == 0 &&
|
|
"chunkWidth must divide width of the big integer");
|
|
auto numberOfChunks = bigIntWidth / chunkWidth;
|
|
std::vector<int64_t> shape({numberOfChunks});
|
|
return mlir::RankedTensorType::get(shape, eint);
|
|
}
|
|
|
|
/// The type converter used to transform `FHE` ops on chunked integers
|
|
class TypeConverter : public mlir::TypeConverter {
|
|
|
|
public:
|
|
TypeConverter(unsigned int chunkSize, unsigned int chunkWidth) {
|
|
addConversion([](mlir::Type type) { return type; });
|
|
addConversion([chunkSize,
|
|
chunkWidth](FHE::ChunkedEncryptedIntegerType type) {
|
|
return convertChunkedEint(type.getContext(), type, chunkSize, chunkWidth);
|
|
});
|
|
}
|
|
};
|
|
|
|
} // namespace typing
|
|
|
|
class AddEintPattern
|
|
: public mlir::OpConversionPattern<mlir::concretelang::FHE::AddEintOp> {
|
|
public:
|
|
AddEintPattern(mlir::TypeConverter &converter, mlir::MLIRContext *context,
|
|
unsigned int chunkSize, unsigned int chunkWidth)
|
|
: mlir::OpConversionPattern<mlir::concretelang::FHE::AddEintOp>(
|
|
converter, context, ::mlir::concretelang::DEFAULT_PATTERN_BENEFIT),
|
|
chunkSize(chunkSize), chunkWidth(chunkWidth) {}
|
|
|
|
mlir::LogicalResult
|
|
matchAndRewrite(FHE::AddEintOp op, FHE::AddEintOp::Adaptor adaptor,
|
|
mlir::ConversionPatternRewriter &rewriter) const override {
|
|
auto tensorType = adaptor.a().getType().dyn_cast<mlir::RankedTensorType>();
|
|
auto shape = tensorType.getShape();
|
|
assert(shape.size() == 1 &&
|
|
"chunked integer should be converted to flat tensors, but tensor "
|
|
"have more than one dimension");
|
|
auto eintChunkWidth = tensorType.getElementType()
|
|
.dyn_cast<FHE::EncryptedIntegerType>()
|
|
.getWidth();
|
|
assert(eintChunkWidth == chunkSize && "wrong tensor elements width");
|
|
auto numberOfChunks = shape[0];
|
|
|
|
mlir::Value carry =
|
|
rewriter
|
|
.create<FHE::ZeroEintOp>(op.getLoc(),
|
|
FHE::EncryptedIntegerType::get(
|
|
rewriter.getContext(), chunkSize))
|
|
.getResult();
|
|
|
|
mlir::Value resultTensor =
|
|
rewriter.create<FHE::ZeroTensorOp>(op.getLoc(), adaptor.a().getType())
|
|
.getResult();
|
|
// used to shift the carry bit to the left
|
|
mlir::Value twoPowerChunkSizeCst =
|
|
rewriter
|
|
.create<mlir::arith::ConstantIntOp>(op.getLoc(), 1 << chunkWidth,
|
|
chunkSize + 1)
|
|
.getResult();
|
|
// Create the loop
|
|
int64_t lb = 0, step = 1;
|
|
auto forOp = rewriter.create<mlir::AffineForOp>(
|
|
op.getLoc(), lb, numberOfChunks, step, resultTensor,
|
|
[&](mlir::OpBuilder &builder, mlir::Location loc, mlir::Value iter,
|
|
mlir::ValueRange args) {
|
|
// add inputs with the previous carry (init to 0)
|
|
mlir::Value leftEint =
|
|
builder.create<mlir::tensor::ExtractOp>(loc, adaptor.a(), iter);
|
|
mlir::Value rightEint =
|
|
builder.create<mlir::tensor::ExtractOp>(loc, adaptor.b(), iter);
|
|
mlir::Value result =
|
|
builder.create<FHE::AddEintOp>(loc, leftEint, rightEint)
|
|
.getResult();
|
|
mlir::Value resultWithCarry =
|
|
builder.create<FHE::AddEintOp>(loc, result, carry).getResult();
|
|
// compute the new carry: either 1 or 0
|
|
carry =
|
|
rewriter.create<mlir::concretelang::FHE::ApplyLookupTableEintOp>(
|
|
op.getLoc(),
|
|
FHE::EncryptedIntegerType::get(rewriter.getContext(),
|
|
chunkSize),
|
|
resultWithCarry,
|
|
getTruthTableCarryExtract(rewriter, op.getLoc(), chunkSize,
|
|
chunkWidth));
|
|
// remove the carry bit from the result
|
|
mlir::Value shiftedCarry =
|
|
builder
|
|
.create<FHE::MulEintIntOp>(loc, carry, twoPowerChunkSizeCst)
|
|
.getResult();
|
|
mlir::Value finalResult =
|
|
builder.create<FHE::SubEintOp>(loc, resultWithCarry, shiftedCarry)
|
|
.getResult();
|
|
// insert the result in the result tensor
|
|
mlir::Value tensorResult = builder.create<mlir::tensor::InsertOp>(
|
|
loc, finalResult, args[0], iter);
|
|
builder.create<mlir::AffineYieldOp>(loc, tensorResult);
|
|
});
|
|
rewriter.replaceOp(op, forOp.getResult(0));
|
|
return mlir::success();
|
|
}
|
|
|
|
private:
|
|
unsigned int chunkSize, chunkWidth;
|
|
};
|
|
|
|
/// Perfoms the transformation of big integer operations
|
|
class FHEBigIntTransformPass
|
|
: public FHEBigIntTransformBase<FHEBigIntTransformPass> {
|
|
public:
|
|
FHEBigIntTransformPass(unsigned int chunkSize, unsigned int chunkWidth)
|
|
: chunkSize(chunkSize), chunkWidth(chunkWidth){};
|
|
|
|
void runOnOperation() override {
|
|
mlir::Operation *op = getOperation();
|
|
mlir::ConversionTarget target(getContext());
|
|
mlir::RewritePatternSet patterns(&getContext());
|
|
typing::TypeConverter converter(chunkSize, chunkWidth);
|
|
|
|
// Legal ops created during pattern application
|
|
target.addLegalOp<mlir::AffineForOp, mlir::AffineYieldOp,
|
|
mlir::arith::ConstantOp, mlir::arith::ConstantIndexOp,
|
|
FHE::ZeroEintOp, FHE::ZeroTensorOp, FHE::AddEintOp,
|
|
FHE::MulEintIntOp, FHE::SubEintOp,
|
|
FHE::ApplyLookupTableEintOp, mlir::tensor::ExtractOp,
|
|
mlir::tensor::InsertOp>();
|
|
concretelang::addDynamicallyLegalTypeOp<FHE::AddEintOp>(target, converter);
|
|
// Func ops are only legal with converted types
|
|
target.addDynamicallyLegalOp<mlir::func::FuncOp>(
|
|
[&](mlir::func::FuncOp funcOp) {
|
|
return converter.isSignatureLegal(funcOp.getFunctionType()) &&
|
|
converter.isLegal(&funcOp.getBody());
|
|
});
|
|
mlir::populateFunctionOpInterfaceTypeConversionPattern<mlir::func::FuncOp>(
|
|
patterns, converter);
|
|
patterns.add<concretelang::TypeConvertingReinstantiationPattern<
|
|
mlir::func::ReturnOp>>(patterns.getContext(), converter);
|
|
concretelang::addDynamicallyLegalTypeOp<mlir::func::ReturnOp>(target,
|
|
converter);
|
|
|
|
patterns.add<AddEintPattern>(converter, &getContext(), chunkSize,
|
|
chunkWidth);
|
|
|
|
if (mlir::applyPartialConversion(op, target, std::move(patterns))
|
|
.failed()) {
|
|
this->signalPassFailure();
|
|
}
|
|
}
|
|
|
|
private:
|
|
unsigned int chunkSize, chunkWidth;
|
|
};
|
|
|
|
} // end anonymous namespace
|
|
|
|
std::unique_ptr<mlir::OperationPass<>>
|
|
createFHEBigIntTransformPass(unsigned int chunkSize, unsigned int chunkWidth) {
|
|
assert(chunkSize >= chunkWidth + 1 &&
|
|
"chunkSize must be greater than chunkWidth");
|
|
return std::make_unique<FHEBigIntTransformPass>(chunkSize, chunkWidth);
|
|
}
|
|
|
|
} // namespace concretelang
|
|
} // namespace mlir
|