* Copyright (c) 2025 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.
*/
* \file codegen_for_block.h
* \brief
*/
#pragma once
#include <unordered_set>
#include <vector>
#include "interface/operation/opcode.h"
#include "interface/tensor/symbolic_scalar.h"
#include "codegen/symbol_mgr/codegen_symbol.h"
namespace npu::tile_fwk {
constexpr const int MAX_LOOP_DEPTH = 3;
struct ForNode {
std::string loopVar;
SymbolicScalar start;
SymbolicScalar extent;
SymbolicScalar step;
std::string Print() const;
void PrintInit(std::ostringstream& os) const;
void PrintCond(std::ostringstream& os) const;
void PrintUpdate(std::ostringstream& os) const;
};
using OffsetInLoop = std::vector<std::string>;
class ForBlockManager {
public:
explicit ForBlockManager(const std::shared_ptr<SymbolManager>& symbolManager) : sm_(symbolManager)
{
forNodes_.reserve(MAX_LOOP_DEPTH);
};
~ForBlockManager() = default;
void UpdateAxesList(const std::vector<SymbolicScalar>& axesList);
void LoopStart() { isInLoop_ = true; }
void OutLoop()
{
sm_->OutForLoop();
axesList_.clear();
forNodes_.clear();
tensorOffset_.clear();
allOffsetsInLoop_.clear();
opList_.clear();
isInLoop_ = false;
offsetCnt_ = 0;
}
bool IsInLoop() { return isInLoop_; }
void AddTensorInLoopBody(
const std::string& tensorFullDim, const TileTensor& tileTensor, const Operation& oper, Opcode opCode);
void AddOpInLoopBody(std::string& op)
{
CODEGEN_LOGI("AddOpInLoopBody add op : %s", op.c_str());
opList_.emplace_back(op);
}
std::string Print() const;
private:
void PrintForHeader(std::ostringstream& os) const;
void PrintForBody(std::ostringstream& os) const;
void PrintForEnd(std::ostringstream& os) const;
void PrintOffsetDef(std::ostringstream& os) const;
void PrintSetAddrs(std::ostringstream& os) const;
void PrintSetAddrSingle(std::ostringstream& os, const std::string& tensor, const std::string& offset) const;
void PrintTileOps(std::ostringstream& os) const;
void UpdateTensorOffsetInLoop(
Opcode opCode, int tensorMagic, const Operation& oper, const std::string& tensorNameInLoop);
OffsetInLoop BuildOffsetInLoop(int tensorMagic, const Operation& oper);
std::shared_ptr<SymbolManager> sm_;
std::vector<SymbolicScalar> axesList_;
std::vector<ForNode> forNodes_;
std::vector<std::string> opList_;
std::unordered_map<std::string, std::string> tensorOffset_;
std::map<OffsetInLoop, std::string> allOffsetsInLoop_;
OffsetInLoop defaultOffset_;
bool isInLoop_{false};
int offsetCnt_{0};
};
}