* 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 kernel_tensor_base.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message("impl/basic_api/kernel_tensor_base.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"basic_api/kernel_tpipe.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_TENSOR_BASE_H__
#endif
#ifndef ASCENDC_MODULE_TENSOR_BASE_H
#define ASCENDC_MODULE_TENSOR_BASE_H
#include <cstdint>
#include "kernel_utils.h"
#include "kernel_operator_layout.h"
#include "kernel_operator_tensor_trait.h"
namespace AscendC {
using TBufHandle = uint8_t*;
using TEventID = int8_t;
using TTagType = int32_t;
template <typename T>
struct GetTypeFromTrait;
template <typename T, TPosition pos, typename Shape, typename Stride>
struct GetTypeFromTrait<TensorTrait<T, pos, Layout<Shape, Stride>>> {
using LayoutType = Layout<Shape, Stride>;
using ShapeType = Shape;
using StrideType = Stride;
using ShapeRows = typename Std::tuple_element<0, Shape>::type;
using ShapeColumns = typename Std::tuple_element<1, Shape>::type;
using StrideRows = typename Std::tuple_element<0, Stride>::type;
using StrideColumns = typename Std::tuple_element<1, Stride>::type;
};
template <typename T, TPosition pos, typename Shape, typename Stride>
struct GetTypeFromTrait<LocalTensor<TensorTrait<T, pos, Layout<Shape, Stride>>>>
: public GetTypeFromTrait<TensorTrait<T, pos, Layout<Shape, Stride>>> {};
template <typename T, TPosition pos, typename Shape, typename Stride>
struct GetTypeFromTrait<GlobalTensor<TensorTrait<T, pos, Layout<Shape, Stride>>>>
: public GetTypeFromTrait<TensorTrait<T, pos, Layout<Shape, Stride>>> {};
template <typename T>
using GetLayoutType = Std::remove_cvref_t<typename GetTypeFromTrait<T>::LayoutType>;
template <typename T>
using GetShapeType = Std::remove_cvref_t<typename GetTypeFromTrait<T>::ShapeType>;
template <typename T>
using GetStrideType = Std::remove_cvref_t<typename GetTypeFromTrait<T>::StrideType>;
template <typename T>
using GetShapeRows = Std::remove_cvref_t<typename GetTypeFromTrait<T>::ShapeRows>;
template <typename T>
using GetShapeColumns = Std::remove_cvref_t<typename GetTypeFromTrait<T>::ShapeColumns>;
template <typename T>
using GetStrideRows = Std::remove_cvref_t<typename GetTypeFromTrait<T>::StrideRows>;
template <typename T>
using GetStrideColumns = Std::remove_cvref_t<typename GetTypeFromTrait<T>::StrideColumns>;
enum class TBufState : uint8_t {
FREE = 0,
OCCUPIED,
ENQUE,
DEQUE,
};
struct TBufType {
TBufState state;
HardEvent freeBufEvt;
#if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
union {
struct {
TEventID enQueEvtID;
TEventID freeBufEvtID;
};
struct {
TBufId bufId;
TBufId bufIdAlt;
};
};
#else
TEventID enQueEvtID;
TEventID freeBufEvtID;
#endif
uint32_t address;
uint32_t dataLen;
TTagType usertag;
DEBUG_CODE(HardEvent userEnQueEvt);
};
struct TBuffAddr {
uint32_t dataLen;
uint32_t bufferAddr;
TBufHandle bufferHandle;
uint8_t logicPos;
#if defined(ASCENDC_CPU_DEBUG) && ASCENDC_CPU_DEBUG == 1
uint8_t* absAddr;
#endif
};
template <typename T> class BaseLocalTensor {
public:
using PrimType = PrimT<T>;
#if defined(ASCENDC_CPU_DEBUG) && ASCENDC_CPU_DEBUG == 1
void SetAddr(const TBuffAddr& address)
{
this->address_ = address;
}
[[deprecated("NOTICE: InitBuffer has been deprecated and will be removed in the next version. "
"Please do not use it!")]]
void InitBuffer(const uint32_t bufferOffset, const uint32_t bufferSize)
{
#if (__NPU_ARCH__ == 5102)
if constexpr (IsSameType<PrimType, int4b_t>::value || IsSameType<PrimType, int2b_t>::value || IsSameType<PrimType, uint1b_t>::value) {
#else
if constexpr (IsSameType<PrimType, int4b_t>::value) {
#endif
ASCENDC_DEBUG_ASSERT((bufferSize != 0),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "InitBuffer bufferSize must be larger than 0."));
ASCENDC_DEBUG_ASSERT((bufferOffset % ONE_BLK_SIZE == 0),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "bufferOffset is %u, which should be 32Bytes aligned", bufferOffset));
}
ASCENDC_DEBUG_ASSERT((TPosition(this->address_.logicPos) != TPosition::GM),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "logicPos can not be gm when init buffer"));
auto positionHardMap = ConstDefiner::Instance().positionHardMap;
ASCENDC_DEBUG_ASSERT((positionHardMap.count(AscendC::TPosition(this->address_.logicPos)) != 0),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "illegal logic pos %d", this->address_.logicPos));
if constexpr (IsSameType<PrimType, int4b_t>::value) {
ASCENDC_DEBUG_ASSERT((bufferOffset + bufferSize * INT4_BIT_NUM / ONE_BYTE_BIT_SIZE <=
ConstDefiner::Instance().bufferInitLen.at(positionHardMap.at(
AscendC::TPosition(this->address_.logicPos)))),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "bufferOffset is %d, bufferSize is %d, buffer overflow",
bufferOffset, bufferSize));
this->address_.absAddr = ConstDefiner::Instance().hardwareCpuBufferMap.at(
positionHardMap.at(AscendC::TPosition(this->address_.logicPos))) +
bufferOffset;
this->address_.dataLen = bufferSize * INT4_BIT_NUM / ONE_BYTE_BIT_SIZE;
#if (__NPU_ARCH__ == 5102)
} else if constexpr (IsSameType<PrimType, int2b_t>::value) {
ASCENDC_DEBUG_ASSERT((bufferOffset + bufferSize * INT2_BIT_NUM / ONE_BYTE_BIT_SIZE <=
ConstDefiner::Instance().bufferInitLen.at(positionHardMap.at(
AscendC::TPosition(this->address_.logicPos)))),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "bufferOffset is %d, bufferSize is %d, buffer overflow",
bufferOffset, bufferSize));
this->address_.absAddr = ConstDefiner::Instance().hardwareCpuBufferMap.at(
positionHardMap.at(AscendC::TPosition(this->address_.logicPos))) +
bufferOffset;
this->address_.dataLen = bufferSize * INT2_BIT_NUM / ONE_BYTE_BIT_SIZE;
} else if constexpr (IsSameType<PrimType, uint1b_t>::value) {
ASCENDC_DEBUG_ASSERT((bufferOffset + bufferSize * INT1_BIT_NUM / ONE_BYTE_BIT_SIZE <=
ConstDefiner::Instance().bufferInitLen.at(positionHardMap.at(
AscendC::TPosition(this->address_.logicPos)))),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "bufferOffset is %d, bufferSize is %d, buffer overflow",
bufferOffset, bufferSize));
this->address_.absAddr = ConstDefiner::Instance().hardwareCpuBufferMap.at(
positionHardMap.at(AscendC::TPosition(this->address_.logicPos))) +
bufferOffset;
this->address_.dataLen = bufferSize * INT1_BIT_NUM / ONE_BYTE_BIT_SIZE;
#endif
} else {
ASCENDC_DEBUG_ASSERT((bufferOffset % ONE_BLK_SIZE == 0),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "bufferOffset is %u, which should be 32Bytes aligned", bufferOffset));
ASCENDC_DEBUG_ASSERT((bufferOffset + bufferSize * sizeof(PrimType) <=
ConstDefiner::Instance().bufferInitLen.at(positionHardMap.at(
AscendC::TPosition(this->address_.logicPos)))),
KERNEL_LOG_INTERNAL(KERNEL_ERROR,
"bufferOffset is %d, bufferSize is %d, buffer overflow",
bufferOffset, bufferSize));
ASCENDC_DEBUG_ASSERT((bufferSize != 0),
KERNEL_LOG_INTERNAL(KERNEL_ERROR, "InitBuffer bufferSize must be larger than 0."));
this->address_.absAddr = ConstDefiner::Instance().hardwareCpuBufferMap.at(
positionHardMap.at(AscendC::TPosition(this->address_.logicPos))) +
bufferOffset;
this->address_.dataLen = bufferSize * sizeof(PrimType);
}
}
#else
__aicore__ inline void SetAddr(const TBuffAddr& bufferAddr)
{
this->address_ = bufferAddr;
}
[[deprecated("NOTICE: InitBuffer has been deprecated and will be removed in the next version. "
"Please do not use it!")]]
__aicore__ inline void InitBuffer(const uint32_t bufferOffset, const uint32_t bufferSize)
{
this->address_.bufferAddr = get_imm(0) + bufferOffset;
if constexpr (IsSameType<PrimType, int4b_t>::value) {
this->address_.dataLen = bufferSize * INT4_BIT_NUM / ONE_BYTE_BIT_SIZE;
#if (__NPU_ARCH__ == 5102)
} else if constexpr (IsSameType<PrimType, int2b_t>::value) {
this->address_.dataLen = bufferSize * INT2_BIT_NUM / ONE_BYTE_BIT_SIZE;
} else if constexpr (IsSameType<PrimType, uint1b_t>::value) {
this->address_.dataLen = bufferSize * INT1_BIT_NUM / ONE_BYTE_BIT_SIZE;
#endif
} else {
this->address_.dataLen = bufferSize * sizeof(PrimType);
}
}
#endif
__aicore__ inline TBufHandle GetBufferHandle() const
{
return address_.bufferHandle;
}
public:
TBuffAddr address_;
};
template <typename T> class BaseGlobalTensor {
public:
using PrimType = PrimT<T>;
__aicore__ inline void SetAddr(const uint64_t offset)
{
if constexpr (IsSameType<PrimType, int4b_t>::value) {
address_ = address_ + offset / INT4_TWO;
oriAddress_ = oriAddress_ + offset / INT4_TWO;
#if (__NPU_ARCH__ == 5102)
} else if constexpr (IsSameType<PrimType, int2b_t>::value) {
address_ = address_ + offset / INT2_FOUR;
oriAddress_ = oriAddress_ + offset / INT2_FOUR;
} else if constexpr (IsSameType<PrimType, uint1b_t>::value) {
address_ = address_ + offset / INT1_EIGHT;
oriAddress_ = oriAddress_ + offset / INT1_EIGHT;
#endif
} else {
address_ = address_ + offset;
oriAddress_ = oriAddress_ + offset;
}
}
public:
__gm__ PrimType* address_;
__gm__ PrimType* oriAddress_;
};
template <typename T> class BaseTensor {};
template <typename T> class BaseTensorTraitTensor {};
template <typename T, TPosition pos, typename LayoutType>
class BaseTensorTraitTensor<TensorTrait<T, pos, LayoutType>> {
public:
__aicore__ inline TensorTrait<T, pos, LayoutType>& GetTensorTrait();
__aicore__ inline const TensorTrait<T, pos, LayoutType>& GetTensorTrait() const;
__aicore__ inline void SetTensorTrait(const TensorTrait<T, pos, LayoutType>& newTrait);
__aicore__ inline decltype(auto) GetLayout() const { return trait.GetLayout(); }
__aicore__ inline decltype(auto) GetShape() const { return trait.GetShape(); }
__aicore__ inline decltype(auto) GetStride() const { return trait.GetStride(); }
private:
TensorTrait<T, pos, LayoutType> trait = {};
};
template <typename T, TPosition pos, typename LayoutType>
__aicore__ inline TensorTrait<T, pos, LayoutType>& BaseTensorTraitTensor<TensorTrait<T, pos, LayoutType>>::GetTensorTrait() {
return this->trait;
}
template <typename T, TPosition pos, typename LayoutType>
__aicore__ inline const TensorTrait<T, pos, LayoutType>& BaseTensorTraitTensor<TensorTrait<T, pos, LayoutType>>::GetTensorTrait() const {
return this->trait;
}
template <typename T, TPosition pos, typename LayoutType>
__aicore__ inline void BaseTensorTraitTensor<TensorTrait<T, pos, LayoutType>>::SetTensorTrait(const TensorTrait<T, pos, LayoutType>& newTrait) {
this->trait = newTrait;
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_TENSOR_BASE_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_TENSOR_BASE_H__
#endif
