* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include "pass_operation_utils.h"
#include <algorithm>
#include "interface/operation/op_infer_shape_impl.h"
#include "interface/operation/operation_common.h"
#include "interface/tensor/irbuilder.h"
#include "interface/tensor/tensor_offset.h"
namespace npu::tile_fwk {
Operation& PassOperationUtils::AddOperation(
Function& function, Opcode opCode, LogicalTensors iOperands, const LogicalTensors& oOperands,
std::function<void(Operation&)> beforeInferShapeHandler, const ir::Span& span, bool inferShape)
{
auto processedOperands = PreprocessOperationInputs(function, opCode, std::move(iOperands));
IRBuilder builder;
auto& op = builder.CreateTensorOpStmt(function, opCode, processedOperands, oOperands, span);
if (beforeInferShapeHandler) {
beforeInferShapeHandler(op);
}
if (inferShape) {
InferShapeRegistry::GetInstance().CallInferShapeFunc(&op);
}
return op;
}
LogicalTensors PassOperationUtils::PreprocessOperationInputs(
Function& function, Opcode opCode, LogicalTensors iOperands)
{
CheckTensorDynamicShape(iOperands, opCode);
for (auto& iOperand : iOperands) {
FE_ASSERT(FeError::INVALID_VAL, iOperand->shape.size() != 0) << "tensor shape size invalid";
iOperand = ConnectWithOverlap(function, iOperand);
}
return iOperands;
}
std::vector<std::vector<int64_t>> PassOperationUtils::ProcessOffsetAdjustment(
LogicalTensors& matches, std::vector<int64_t>& minimumOffsets)
{
std::vector<std::vector<int64_t>> offsetOfOverlaps;
std::sort(matches.begin(), matches.end(), [](const auto& a, const auto& b) { return a->offset < b->offset; });
minimumOffsets = matches.front()->offset;
for (auto& match : matches) {
offsetOfOverlaps.emplace_back(match->offset);
for (size_t idx = 0; idx < minimumOffsets.size(); ++idx) {
minimumOffsets[idx] = std::min(minimumOffsets[idx], match->offset[idx]);
}
}
for (auto& offsetOfOverlap : offsetOfOverlaps) {
for (size_t idx = 0; idx < offsetOfOverlap.size(); ++idx) {
offsetOfOverlap[idx] -= minimumOffsets[idx];
}
}
return offsetOfOverlaps;
}
LogicalTensorPtr PassOperationUtils::HandlePerfectlyMatchWithAll(
Function& function, LogicalTensorPtr iOperand, const LogicalTensors& matches,
const std::vector<std::vector<int64_t>>& offsetOfOverlaps)
{
auto assembleResult = std::make_shared<LogicalTensor>(
function, iOperand->Datatype(), iOperand->shape, iOperand->GetDynValidShape(), iOperand->Format(),
"Assemble_" + matches[0]->Symbol());
FE_ASSERT(FeError::NOT_EXIST, assembleResult->GetProducers().empty()) << "Assemble result should have no producers";
IRBuilder builder;
for (size_t idx = 0; idx < matches.size(); ++idx) {
auto& assembleOp = builder.CreateTensorOpStmt(function, Opcode::OP_ASSEMBLE, {matches[idx]}, {assembleResult});
assembleOp.SetOpAttribute(std::make_shared<AssembleOpAttribute>(
offsetOfOverlaps[idx], SymbolicScalar::FromConcrete(offsetOfOverlaps[idx])));
}
return assembleResult;
}
LogicalTensorPtr PassOperationUtils::HandleBeCovered(
Function& function, LogicalTensorPtr iOperand, const LogicalTensors& matches)
{
auto viewResult = std::make_shared<LogicalTensor>(
function, matches.front()->tensor->datatype, iOperand->shape, iOperand->Format(),
"View_" + matches.front()->tensor->symbol);
IRBuilder irBuilder;
auto& viewOp = irBuilder.CreateTensorOpStmt(function, Opcode::OP_VIEW, {matches.front()}, {viewResult});
viewOp.SetOpAttribute(std::make_shared<ViewOpAttribute>(
iOperand->GetOffset(), iOperand->GetDynOffset(), iOperand->GetDynValidShape()));
if (!iOperand->GetDynValidShape().empty()) {
viewResult->UpdateDynValidShape(iOperand->GetDynValidShape());
}
return viewResult;
}
LogicalTensorPtr PassOperationUtils::HandleBeCoveredByAll(
Function& function, LogicalTensorPtr iOperand, const LogicalTensors& matches,
const std::vector<std::vector<int64_t>>& offsetOfOverlaps)
{
std::vector<int64_t> minimumOffset;
std::vector<int64_t> maximumShape;
CalcShapeAndOffsetOfGroup(matches, minimumOffset, maximumShape);
auto assembleResult = std::make_shared<LogicalTensor>(
function, matches[0]->Datatype(), maximumShape, iOperand->Format(),
"Assemble_" + matches[0]->Symbol());
FE_ASSERT(FeError::NOT_EXIST, assembleResult->GetProducers().empty()) << "Assemble result should have no producers";
IRBuilder builder;
for (size_t idx = 0; idx < matches.size(); ++idx) {
auto& assembleOp = builder.CreateTensorOpStmt(function, Opcode::OP_ASSEMBLE, {matches[idx]}, {assembleResult});
assembleOp.SetOpAttribute(std::make_shared<AssembleOpAttribute>(
offsetOfOverlaps[idx], SymbolicScalar::FromConcrete(offsetOfOverlaps[idx])));
}
auto viewResult = std::make_shared<LogicalTensor>(
function, assembleResult->Datatype(), iOperand->shape, iOperand->Format(),
"View_" + assembleResult->Symbol());
auto& viewOp = builder.CreateTensorOpStmt(function, Opcode::OP_VIEW, {assembleResult}, {viewResult});
std::vector<int64_t> newOffset = TensorOffset::Sub(iOperand->GetOffset(), minimumOffset);
std::vector<SymbolicScalar> newDynOffset = TensorOffset::Sub(iOperand->GetDynOffset(), minimumOffset);
viewOp.SetOpAttribute(std::make_shared<ViewOpAttribute>(newOffset, newDynOffset, iOperand->GetDynValidShape()));
return viewResult;
}
LogicalTensorPtr PassOperationUtils::ConnectWithOverlap(Function& function, LogicalTensorPtr iOperand)
{
auto matches = function.GetTensorMap().Find(iOperand);
if (matches.empty()) {
return iOperand;
}
auto overlapStatus = CalcOverlap(iOperand, matches);
FE_ASSERT(!matches.empty()) << "Matches should not be empty";
std::vector<int64_t> minimumOffsets;
auto offsetOfOverlaps = ProcessOffsetAdjustment(matches, minimumOffsets);
switch (overlapStatus) {
case OverlapStatus::PERFECTLY_MATCH_WITH_ALL:
return HandlePerfectlyMatchWithAll(function, iOperand, matches, offsetOfOverlaps);
case OverlapStatus::PERFECTLY_MATCH:
return matches.front();
case OverlapStatus::BE_COVERED:
return HandleBeCovered(function, iOperand, matches);
case OverlapStatus::BE_COVERED_BY_ALL:
return HandleBeCoveredByAll(function, iOperand, matches, offsetOfOverlaps);
default:
FE_ASSERT(false) << "unexpected behavior";
}
FE_ASSERT(false) << "unexpected behavior";
return nullptr;
}
}