Files
concrete/compiler/lib/Dialect/BConcrete/Transforms/BufferizableOpInterfaceImpl.cpp
2022-10-07 09:16:19 +02:00

605 lines
24 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/Arithmetic/IR/Arithmetic.h"
#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
#include "mlir/Dialect/Bufferization/Transforms/BufferUtils.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/IR/Dialect.h"
#include "mlir/IR/Operation.h"
#include "concretelang/Conversion/Tools.h"
#include "concretelang/Dialect/BConcrete/IR/BConcreteDialect.h"
#include "concretelang/Dialect/BConcrete/IR/BConcreteOps.h"
#include "concretelang/Dialect/BConcrete/Transforms/BufferizableOpInterfaceImpl.h"
#include <mlir/IR/AffineExpr.h>
#include <mlir/IR/AffineMap.h>
#include <mlir/IR/BuiltinTypes.h>
using namespace mlir;
using namespace mlir::bufferization;
using namespace mlir::tensor;
namespace BConcrete = mlir::concretelang::BConcrete;
namespace mlir {
namespace concretelang {
namespace BConcrete {
namespace {} // namespace
} // namespace BConcrete
} // namespace concretelang
} // namespace mlir
namespace {
mlir::Type getDynamicMemrefWithUnknownOffset(mlir::RewriterBase &rewriter,
size_t rank) {
std::vector<int64_t> shape(rank, -1);
mlir::AffineExpr expr = rewriter.getAffineSymbolExpr(0);
for (size_t i = 0; i < rank; i++) {
expr = expr +
(rewriter.getAffineDimExpr(i) * rewriter.getAffineSymbolExpr(i + 1));
}
return mlir::MemRefType::get(
shape, rewriter.getI64Type(),
mlir::AffineMap::get(rank, rank + 1, expr, rewriter.getContext()));
}
// Returns `memref.cast %0 : memref<...xAxT> to memref<...x?xT>`
mlir::Value getCastedMemRef(mlir::RewriterBase &rewriter, mlir::Location loc,
mlir::Value value) {
mlir::Type valueType = value.getType();
if (auto memrefTy = valueType.dyn_cast_or_null<mlir::MemRefType>()) {
return rewriter.create<mlir::memref::CastOp>(
loc,
getDynamicMemrefWithUnknownOffset(rewriter, memrefTy.getShape().size()),
value);
} else {
return value;
}
}
char memref_add_lwe_ciphertexts_u64[] = "memref_add_lwe_ciphertexts_u64";
char memref_add_plaintext_lwe_ciphertext_u64[] =
"memref_add_plaintext_lwe_ciphertext_u64";
char memref_mul_cleartext_lwe_ciphertext_u64[] =
"memref_mul_cleartext_lwe_ciphertext_u64";
char memref_negate_lwe_ciphertext_u64[] = "memref_negate_lwe_ciphertext_u64";
char memref_keyswitch_lwe_u64[] = "memref_keyswitch_lwe_u64";
char memref_bootstrap_lwe_u64[] = "memref_bootstrap_lwe_u64";
char memref_keyswitch_async_lwe_u64[] = "memref_keyswitch_async_lwe_u64";
char memref_bootstrap_async_lwe_u64[] = "memref_bootstrap_async_lwe_u64";
char memref_await_future[] = "memref_await_future";
char memref_expand_lut_in_trivial_glwe_ct_u64[] =
"memref_expand_lut_in_trivial_glwe_ct_u64";
char memref_wop_pbs_crt_buffer[] = "memref_wop_pbs_crt_buffer";
mlir::LogicalResult insertForwardDeclarationOfTheCAPI(
mlir::Operation *op, mlir::RewriterBase &rewriter, char const *funcName) {
auto memref1DType = getDynamicMemrefWithUnknownOffset(rewriter, 1);
auto memref2DType = getDynamicMemrefWithUnknownOffset(rewriter, 2);
auto futureType =
mlir::concretelang::RT::FutureType::get(rewriter.getIndexType());
auto contextType =
mlir::concretelang::Concrete::ContextType::get(rewriter.getContext());
mlir::FunctionType funcType;
if (funcName == memref_add_lwe_ciphertexts_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(), {memref1DType, memref1DType, memref1DType}, {});
} else if (funcName == memref_add_plaintext_lwe_ciphertext_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(),
{memref1DType, memref1DType, rewriter.getI64Type()}, {});
} else if (funcName == memref_mul_cleartext_lwe_ciphertext_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(),
{memref1DType, memref1DType, rewriter.getI64Type()}, {});
} else if (funcName == memref_negate_lwe_ciphertext_u64) {
funcType = mlir::FunctionType::get(rewriter.getContext(),
{memref1DType, memref1DType}, {});
} else if (funcName == memref_keyswitch_lwe_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(), {memref1DType, memref1DType, contextType}, {});
} else if (funcName == memref_bootstrap_lwe_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(),
{memref1DType, memref1DType, memref1DType, contextType}, {});
} else if (funcName == memref_keyswitch_async_lwe_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(), {memref1DType, memref1DType, contextType},
{futureType});
} else if (funcName == memref_bootstrap_async_lwe_u64) {
funcType = mlir::FunctionType::get(
rewriter.getContext(),
{memref1DType, memref1DType, memref1DType, contextType}, {futureType});
} else if (funcName == memref_await_future) {
funcType = mlir::FunctionType::get(
rewriter.getContext(),
{memref1DType, futureType, memref1DType, memref1DType}, {});
} else if (funcName == memref_expand_lut_in_trivial_glwe_ct_u64) {
funcType = mlir::FunctionType::get(rewriter.getContext(),
{
memref1DType,
rewriter.getI32Type(),
rewriter.getI32Type(),
rewriter.getI32Type(),
memref1DType,
},
{});
} else if (funcName == memref_wop_pbs_crt_buffer) {
funcType = mlir::FunctionType::get(rewriter.getContext(),
{
memref2DType,
memref2DType,
memref1DType,
memref1DType,
rewriter.getI32Type(),
rewriter.getI32Type(),
rewriter.getI32Type(),
rewriter.getI32Type(),
contextType,
},
{});
} else {
op->emitError("unknwon external function") << funcName;
return mlir::failure();
}
return insertForwardDeclaration(op, rewriter, funcName, funcType);
}
/// Returns the value of the context argument from the enclosing func
mlir::Value getContextArgument(mlir::Operation *op) {
mlir::Block *block = op->getBlock();
while (block != nullptr) {
if (llvm::isa<mlir::func::FuncOp>(block->getParentOp())) {
block = &mlir::cast<mlir::func::FuncOp>(block->getParentOp())
.getBody()
.front();
auto context =
std::find_if(block->getArguments().rbegin(),
block->getArguments().rend(), [](BlockArgument &arg) {
return arg.getType()
.isa<mlir::concretelang::Concrete::ContextType>();
});
assert(context != block->getArguments().rend() &&
"Cannot find the Concrete.context");
return *context;
}
block = block->getParentOp()->getBlock();
}
assert("can't find a function that enclose the op");
return nullptr;
};
template <typename Op>
void pushAdditionalArgs(Op op, mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter);
template <typename Op, char const *funcName>
struct BufferizableWithCallOpInterface
: public BufferizableOpInterface::ExternalModel<
BufferizableWithCallOpInterface<Op, funcName>, Op> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
BufferRelation bufferRelation(Operation *op, OpResult opResult,
const AnalysisState &state) const {
return BufferRelation::None;
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto loc = op->getLoc();
auto castOp = cast<Op>(op);
// For now we always alloc for the result, we didn't have the in place
// operators yet.
auto resTensorType =
castOp.result().getType().template cast<mlir::TensorType>();
auto outMemrefType = MemRefType::get(resTensorType.getShape(),
resTensorType.getElementType());
auto outMemref = options.createAlloc(rewriter, loc, outMemrefType, {});
if (mlir::failed(outMemref)) {
return mlir::failure();
}
// The first operand is the result
mlir::SmallVector<mlir::Value> operands{
getCastedMemRef(rewriter, loc, *outMemref),
};
// For all tensor operand get the corresponding casted buffer
for (auto &operand : op->getOpOperands()) {
if (!operand.get().getType().isa<mlir::RankedTensorType>()) {
operands.push_back(operand.get());
} else {
auto memrefOperand =
bufferization::getBuffer(rewriter, operand.get(), options);
operands.push_back(getCastedMemRef(rewriter, loc, memrefOperand));
}
}
// Append additional argument
pushAdditionalArgs<Op>(castOp, operands, rewriter);
// Insert forward declaration of the function
if (insertForwardDeclarationOfTheCAPI(op, rewriter, funcName).failed()) {
return mlir::failure();
}
rewriter.create<mlir::func::CallOp>(loc, funcName, mlir::TypeRange{},
operands);
replaceOpWithBufferizedValues(rewriter, op, *outMemref);
return success();
}
};
struct BufferizableGlweFromTableOpInterface
: public BufferizableOpInterface::ExternalModel<
BufferizableGlweFromTableOpInterface, BConcrete::FillGlweFromTable> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
BufferRelation bufferRelation(Operation *op, OpResult opResult,
const AnalysisState &state) const {
return BufferRelation::None;
}
/// Bufferize GlweFromTable
/// ```
/// "BConcrete.fill_glwe_table"(%glwe, %lut) {glweDimension=1,
/// polynomialSize=2048, outPrecision=3} :
/// (tensor<4096xi64>, tensor<32xi64>) -> ()
/// ```
///
/// to
///
/// ```
/// %glweDim = arith.constant 1 : i32
/// %polySize = arith.constant 2048 : i32
/// %outPrecision = arith.constant 3 : i32
/// %glwe_ = memref.cast %glwe : memref<4096xi64> to memref<?xi64>
/// %lut_ = memref.cast %lut : memref<32xi64> to memref<?xi64>
/// call @expand_lut_in_trivial_glwe_ct(%glwe, %polySize, %glweDim,
/// %outPrecision, %lut_) :
/// (tensor<?xi64>, i32, i32, tensor<?xi64>) -> ()
/// ```
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto loc = op->getLoc();
auto castOp = cast<BConcrete::FillGlweFromTable>(op);
auto glweOp =
getCastedMemRef(rewriter, loc,
bufferization::getBuffer(
rewriter, castOp->getOpOperand(0).get(), options));
auto lutOp =
getCastedMemRef(rewriter, loc,
bufferization::getBuffer(
rewriter, castOp->getOpOperand(1).get(), options));
auto polySizeOp = rewriter.create<mlir::arith::ConstantOp>(
op->getLoc(), rewriter.getI32IntegerAttr(castOp.polynomialSize()));
auto glweDimensionOp = rewriter.create<mlir::arith::ConstantOp>(
op->getLoc(), rewriter.getI32IntegerAttr(castOp.glweDimension()));
auto outPrecisionOp = rewriter.create<mlir::arith::ConstantOp>(
op->getLoc(), rewriter.getI32IntegerAttr(castOp.outPrecision()));
mlir::SmallVector<mlir::Value> operands{glweOp, polySizeOp, glweDimensionOp,
outPrecisionOp, lutOp};
// Insert forward declaration of the function
if (insertForwardDeclarationOfTheCAPI(
op, rewriter, memref_expand_lut_in_trivial_glwe_ct_u64)
.failed()) {
return mlir::failure();
}
rewriter.create<mlir::func::CallOp>(
loc, memref_expand_lut_in_trivial_glwe_ct_u64, mlir::TypeRange{},
operands);
replaceOpWithBufferizedValues(rewriter, op, {});
return success();
}
};
template <typename Op, char const *funcName, bool withContext = false>
struct BufferizableWithAsyncCallOpInterface
: public BufferizableOpInterface::ExternalModel<
BufferizableWithAsyncCallOpInterface<Op, funcName, withContext>, Op> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
BufferRelation bufferRelation(Operation *op, OpResult opResult,
const AnalysisState &state) const {
return BufferRelation::None;
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto loc = op->getLoc();
auto castOp = cast<Op>(op);
// For now we always alloc for the result, we didn't have the in place
// operators yet.
auto resTensorType =
castOp.result()
.getType()
.template cast<mlir::concretelang::RT::FutureType>()
.getElementType()
.template cast<mlir::TensorType>();
auto outMemrefType = MemRefType::get(resTensorType.getShape(),
resTensorType.getElementType());
auto outMemref = options.createAlloc(rewriter, loc, outMemrefType, {});
if (mlir::failed(outMemref)) {
return mlir::failure();
}
// The first operand is the result
mlir::SmallVector<mlir::Value, 3> operands{
getCastedMemRef(rewriter, loc, *outMemref),
};
// For all tensor operand get the corresponding casted buffer
for (auto &operand : op->getOpOperands()) {
if (!operand.get().getType().isa<mlir::RankedTensorType>()) {
operands.push_back(operand.get());
} else {
auto memrefOperand =
bufferization::getBuffer(rewriter, operand.get(), options);
operands.push_back(getCastedMemRef(rewriter, loc, memrefOperand));
}
}
// Append the context argument
if (withContext) {
operands.push_back(getContextArgument(op));
}
// Insert forward declaration of the function
if (insertForwardDeclarationOfTheCAPI(op, rewriter, funcName).failed()) {
return mlir::failure();
}
auto result = rewriter.create<mlir::func::CallOp>(
loc, funcName,
mlir::TypeRange{
mlir::concretelang::RT::FutureType::get(rewriter.getIndexType())},
operands);
replaceOpWithBufferizedValues(rewriter, op, result.getResult(0));
return success();
}
};
template <typename Op, char const *funcName>
struct BufferizableWithSyncCallOpInterface
: public BufferizableOpInterface::ExternalModel<
BufferizableWithSyncCallOpInterface<Op, funcName>, Op> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
BufferRelation bufferRelation(Operation *op, OpResult opResult,
const AnalysisState &state) const {
return BufferRelation::None;
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto loc = op->getLoc();
auto castOp = cast<Op>(op);
auto resTensorType =
castOp.result().getType().template cast<mlir::TensorType>();
auto outMemrefType = MemRefType::get(resTensorType.getShape(),
resTensorType.getElementType());
auto outMemref = options.createAlloc(rewriter, loc, outMemrefType, {});
if (mlir::failed(outMemref)) {
return mlir::failure();
}
// The first operand is the result
mlir::SmallVector<mlir::Value, 3> operands{
getCastedMemRef(rewriter, loc, *outMemref),
};
// Then add the future operand
operands.push_back(op->getOpOperand(0).get());
// Finally add a dependence on the memref covered by the future to
// prevent early deallocation
auto def = op->getOpOperand(0).get().getDefiningOp();
operands.push_back(def->getOpOperand(0).get());
operands.push_back(def->getOpOperand(1).get());
// Insert forward declaration of the function
if (insertForwardDeclarationOfTheCAPI(op, rewriter, funcName).failed()) {
return mlir::failure();
}
rewriter.create<mlir::func::CallOp>(loc, funcName, mlir::TypeRange{},
operands);
replaceOpWithBufferizedValues(rewriter, op, *outMemref);
return success();
}
};
template <>
void pushAdditionalArgs(BConcrete::AddPlaintextLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter){};
template <>
void pushAdditionalArgs(BConcrete::AddLweBuffersOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter){};
template <>
void pushAdditionalArgs(BConcrete::MulCleartextLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter){};
template <>
void pushAdditionalArgs(BConcrete::NegateLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter){};
template <>
void pushAdditionalArgs(BConcrete::KeySwitchLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter) {
// context
operands.push_back(getContextArgument(op));
};
template <>
void pushAdditionalArgs(BConcrete::BootstrapLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter) {
// context
operands.push_back(getContextArgument(op));
};
template <>
void pushAdditionalArgs(BConcrete::WopPBSCRTLweBufferOp op,
mlir::SmallVector<mlir::Value> &operands,
RewriterBase &rewriter) {
mlir::Type crtType = mlir::RankedTensorType::get(
{(int)op.crtDecompositionAttr().size()}, rewriter.getI64Type());
std::vector<int64_t> values;
for (auto a : op.crtDecomposition()) {
values.push_back(a.cast<IntegerAttr>().getValue().getZExtValue());
}
auto attr = rewriter.getI64TensorAttr(values);
auto x = rewriter.create<mlir::arith::ConstantOp>(op.getLoc(), attr, crtType);
auto globalMemref = bufferization::getGlobalFor(x, 0);
assert(!failed(globalMemref));
auto globalRef = rewriter.create<memref::GetGlobalOp>(
op.getLoc(), (*globalMemref).type(), (*globalMemref).getName());
operands.push_back(getCastedMemRef(rewriter, op.getLoc(), globalRef));
// lwe_small_size
operands.push_back(rewriter.create<mlir::arith::ConstantOp>(
op.getLoc(), op.packingKeySwitchInputLweDimensionAttr()));
// cbs_level_count
operands.push_back(rewriter.create<mlir::arith::ConstantOp>(
op.getLoc(), op.circuitBootstrapLevelAttr()));
// cbs_base_log
operands.push_back(rewriter.create<mlir::arith::ConstantOp>(
op.getLoc(), op.circuitBootstrapBaseLogAttr()));
// polynomial_size
operands.push_back(rewriter.create<mlir::arith::ConstantOp>(
op.getLoc(), op.packingKeySwitchoutputPolynomialSizeAttr()));
// context
operands.push_back(getContextArgument(op));
};
} // namespace
void mlir::concretelang::BConcrete::
registerBufferizableOpInterfaceExternalModels(DialectRegistry &registry) {
registry.addExtension(+[](MLIRContext *ctx,
BConcrete::BConcreteDialect *dialect) {
BConcrete::AddLweBuffersOp::attachInterface<BufferizableWithCallOpInterface<
BConcrete::AddLweBuffersOp, memref_add_lwe_ciphertexts_u64>>(*ctx);
BConcrete::AddPlaintextLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<
BConcrete::AddPlaintextLweBufferOp,
memref_add_plaintext_lwe_ciphertext_u64>>(*ctx);
BConcrete::MulCleartextLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<
BConcrete::MulCleartextLweBufferOp,
memref_mul_cleartext_lwe_ciphertext_u64>>(*ctx);
BConcrete::NegateLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<BConcrete::NegateLweBufferOp,
memref_negate_lwe_ciphertext_u64>>(
*ctx);
BConcrete::KeySwitchLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<BConcrete::KeySwitchLweBufferOp,
memref_keyswitch_lwe_u64>>(*ctx);
BConcrete::BootstrapLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<BConcrete::BootstrapLweBufferOp,
memref_bootstrap_lwe_u64>>(*ctx);
BConcrete::WopPBSCRTLweBufferOp::attachInterface<
BufferizableWithCallOpInterface<BConcrete::WopPBSCRTLweBufferOp,
memref_wop_pbs_crt_buffer>>(*ctx);
BConcrete::KeySwitchLweBufferAsyncOffloadOp::attachInterface<
BufferizableWithAsyncCallOpInterface<
BConcrete::KeySwitchLweBufferAsyncOffloadOp,
memref_keyswitch_async_lwe_u64, true>>(*ctx);
BConcrete::BootstrapLweBufferAsyncOffloadOp::attachInterface<
BufferizableWithAsyncCallOpInterface<
BConcrete::BootstrapLweBufferAsyncOffloadOp,
memref_bootstrap_async_lwe_u64, true>>(*ctx);
BConcrete::AwaitFutureOp::attachInterface<
BufferizableWithSyncCallOpInterface<BConcrete::AwaitFutureOp,
memref_await_future>>(*ctx);
BConcrete::FillGlweFromTable::attachInterface<
BufferizableGlweFromTableOpInterface>(*ctx);
});
}