* 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 distributed_expand.h
* \brief
*/
#ifndef DISTRIBUTED_EXPAND_H
#define DISTRIBUTED_EXPAND_H
#include <cstdint>
#include <vector>
#include <memory>
#include "interface/configs/config_manager.h"
#include "interface/function/function.h"
#include "interface/operation/operation.h"
#include "tilefwk/error_code.h"
namespace npu::tile_fwk {
namespace Distributed {
inline void CreateTileOp(
const TileShape& tileShape, const std::function<void(int32_t, int32_t, int32_t, int32_t, int32_t)>& callback)
{
const auto& tileRow = tileShape.GetDistTileRow();
const auto& tileCol = tileShape.GetDistTileCol();
int32_t rowCount = tileRow[1] + (tileRow[2] == 0 ? 0 : 1);
int32_t colCount = tileCol[1] + (tileCol[2] == 0 ? 0 : 1);
CHECK(DistributedErrorCode::INVALID_TILE_DIM, tileRow[0] > 0)
<< "Invalid tiling strategy of the row axis: the first number must be greater than 0, but got "
<< tileRow[0];
CHECK(DistributedErrorCode::INVALID_TILE_DIM, (tileCol[0] > 0))
<< "Invalid tiling strategy of the col axis: the first number must be greater than 0, but got "
<< tileCol[0];
int32_t tileIndex = 0;
for (int32_t rowIndex = 0; rowIndex < rowCount; rowIndex++) {
int32_t rowShape = ((tileRow[2] != 0) && (rowIndex == rowCount - 1)) ? tileRow[2] : tileRow[0];
for (int32_t colIndex = 0; colIndex < colCount; colIndex++) {
int32_t colShape = ((tileCol[2] != 0) && (colIndex == colCount - 1)) ? tileCol[2] : tileCol[0];
callback(tileIndex, rowIndex * tileRow[0], colIndex * tileCol[0], rowShape, colShape);
tileIndex++;
}
}
}
Tensor SendToRoutingExpert(
const Tensor& shmemData, const Tensor& tokenTensor, const Tensor& tokenExpertTable, const char* group,
const MoeConfig& moeConfig);
void SendToSharedExpert(
const Tensor& shmemData, const Tensor& tokenTensor, const Tensor& syncTensor, const char* group);
Tensor CopyToLocalExpert(const Tensor& tokenTensor, const Tensor& syncTensor);
Tensor DispatchSetFlag(
Tensor& shmemFlag, const Tensor& tokenExpertTable, const Tensor& syncTensor, const char* group,
const MoeConfig& moeConfig);
void TiledSendToRoutingExpert(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledSendToSharedExpert(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledCopyToLocalExpert(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledDispatchSetFlag(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledDispatchFFNSched(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledDispatchFFNBatching(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledDispatchFFNCombineInfo(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledDispatchFFNValidCnt(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemPut(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemStore(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemGet(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemLoad(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemSignal(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemWaitUntil(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemSet(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledShmemBindTensor(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledMoeDistributedCombineSend(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
void TiledMoeDistributedCombineReceive(
Function& function, const TileShape& tileShape, const std::vector<std::shared_ptr<LogicalTensor>>& iOperand,
const std::vector<std::shared_ptr<LogicalTensor>>& oOperand, const Operation& op);
}
}
#endif