* 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_operator_vec_gather_intf_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message("impl/basic_api/kernel_operator_vec_gather_intf_impl.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_operator_vec_gather_intf.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_VEC_GATHER_INTF_IMPL_H__
#endif
#ifndef ASCENDC_MODULE_OPERATOR_VEC_GATHER_INTERFACE_IMPL_H
#define ASCENDC_MODULE_OPERATOR_VEC_GATHER_INTERFACE_IMPL_H
#include "kernel_tensor.h"
#include "kernel_check.h"
#include "kernel_struct_gather.h"
#include "kernel_npu_debug.h"
#if __NPU_ARCH__ == 1001
#include "dav_c100/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 2002
#include "dav_m200/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 2201
#include "dav_c220/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 3002
#include "dav_m300/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 3510
#include "dav_3510/kernel_operator_vec_gather_impl.h"
#elif (__NPU_ARCH__ == 5102)
#include "dav_m510/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 3003
#include "dav_l300/kernel_operator_vec_gather_impl.h"
#elif __NPU_ARCH__ == 3113
#include "dav_l311/kernel_operator_vec_gather_impl.h"
#endif
#pragma begin_pipe(V)
namespace AscendC {
* @ingroup gatherb Level 0
* @brief this function fetches N addresses from offset,then accesses these N addresses(plus the src address)
* @brief to get N 32Byte block, and finally writes these N blocks into dst.
* @brief gather element in the uint of block
* @param [out] dst output LocalTensor
* @param [in] src input LocalTensor
* @param [in] offset input LocalTensor
* @param [in] repeatTime repeat times
* @param [in] repeatParams.dstBlkStride dst block stride
* @param [in] repeatParams.dstRepStride dst repeat stride
*/
template <typename T>
__aicore__ inline void Gatherb(const LocalTensor<T>& dst, const LocalTensor<T>& src,
const LocalTensor<uint32_t>& offset, const uint8_t repeatTime, const GatherRepeatParams& repeatParams)
{
using PrimType = PrimT<T>;
#if ASCENDC_CPU_DEBUG
if (!CheckFuncGatherb(dst, src, offset, repeatTime, repeatParams, "Gatherb")) {
ASCENDC_REPORT_CHECK_ERROR("Gatherb", KernelFuncType::NONE_MODE);
}
#endif
uint32_t srcLength = src.GetSize();
#if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3003) || \
(__NPU_ARCH__ == 3113))
GatherImpl((__ubuf__ PrimType*)dst.GetPhyAddr(), (__ubuf__ PrimType*)src.GetPhyAddr(),
(__ubuf__ uint32_t*)offset.GetPhyAddr(), srcLength, repeatTime, repeatParams);
#else
GatherbImpl((__ubuf__ PrimType*)dst.GetPhyAddr(), (__ubuf__ PrimType*)src.GetPhyAddr(),
(__ubuf__ uint32_t*)offset.GetPhyAddr(), srcLength, repeatTime, repeatParams);
#endif
}
* @ingroup gather Level 0
* @brief gather element from src according to srcOffset
* @param [out] dst output LocalTensor
* @param [in] src input LocalTensor
* @param [in] srcOffset input LocalTensor
* @param [in] srcBaseOffset base address of src
* @param [in] mask valid element count
* @param [in] repeatTime repeat times
* @param [in] dstRepStride dst repeat stride
*/
template <typename T>
__aicore__ inline void Gather(const LocalTensor<T>& dst, const LocalTensor<T>& src,
const LocalTensor<uint32_t>& srcOffset, const uint32_t srcBaseOffset, const uint64_t mask, const uint8_t repeatTime,
const uint16_t dstRepStride)
{
using PrimType = PrimT<T>;
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
CheckVectorTensor("Gather", NamedTensor(dst, "dst"), NamedTensor(src, "src"), NamedTensor(srcOffset, "srcOffset"));
CheckMaskRepeat<PrimType>(mask, repeatTime, "Gather");
CheckValueRange<uint32_t>(srcBaseOffset, 0, src.GetSize() * sizeof(PrimType) - 1, "srcBaseOffset", "Gather");
ASCENDC_DEBUG_ASSERT((srcBaseOffset % sizeof(PrimType) == 0), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check "
"srcBaseOffset value in Gather, it must be divisible by sizeof(T), current srcBaseOffset is %u, sizeof(T) is %u.\n",
srcBaseOffset, sizeof(PrimType)));
#endif
#if ASCENDC_CPU_DEBUG
if (!CheckFuncGather(dst, src, srcOffset, srcBaseOffset, mask, repeatTime, dstRepStride, "Gather")) {
ASCENDC_REPORT_CHECK_ERROR("Gather", KernelFuncType::MASK_COUNT_MODE);
}
#endif
const uint32_t srcLength = src.GetSize();
GatherImpl((__ubuf__ PrimType *)dst.GetPhyAddr(), (__ubuf__ PrimType *)src.GetPhyAddr(),
(__ubuf__ uint32_t *)srcOffset.GetPhyAddr(), srcLength, srcBaseOffset, mask, repeatTime, dstRepStride);
}
* @ingroup gather Level 0
* @brief gather element from src according to srcOffset
* @param [out] dst output LocalTensor
* @param [in] src input LocalTensor
* @param [in] srcOffset input LocalTensor
* @param [in] srcBaseOffset base address of src
* @param [in] mask valid element count(bit mode)
* @param [in] repeatTime repeat times
* @param [in] dstRepStride dst repeat stride
*/
template <typename T>
__aicore__ inline void Gather(const LocalTensor<T>& dst, const LocalTensor<T>& src,
const LocalTensor<uint32_t>& srcOffset, const uint32_t srcBaseOffset, const uint64_t mask[], const uint8_t repeatTime,
const uint16_t dstRepStride)
{
using PrimType = PrimT<T>;
#if defined(ASCENDC_DEBUG) || defined(ASCENDC_CPU_DEBUG)
CheckVectorTensor("Gather", NamedTensor(dst, "dst"), NamedTensor(src, "src"), NamedTensor(srcOffset, "srcOffset"));
CheckMaskRepeat<PrimType>(mask, repeatTime, "Gather");
CheckValueRange<uint32_t>(srcBaseOffset, 0, src.GetSize() * sizeof(PrimType) - 1, "srcBaseOffset", "Gather");
ASCENDC_DEBUG_ASSERT((srcBaseOffset % sizeof(PrimType) == 0), KERNEL_LOG_INTERNAL(KERNEL_ERROR, "Failed to check "
"srcBaseOffset value in Gather, it must be divisible by sizeof(T), current srcBaseOffset is %u, sizeof(T) is %u.\n",
srcBaseOffset, sizeof(PrimType)));
#endif
#if ASCENDC_CPU_DEBUG
if (!CheckFuncGather(dst, src, srcOffset, srcBaseOffset, mask, repeatTime, dstRepStride, "Gather")) {
ASCENDC_REPORT_CHECK_ERROR("Gather", KernelFuncType::MASK_BIT_MODE);
}
#endif
const uint32_t srcLength = src.GetSize();
GatherImpl((__ubuf__ PrimType *)dst.GetPhyAddr(), (__ubuf__ PrimType *)src.GetPhyAddr(),
(__ubuf__ uint32_t *)srcOffset.GetPhyAddr(), srcLength, srcBaseOffset, mask, repeatTime, dstRepStride);
}
* @ingroup gather Level 2
* @brief gather element from src according to srcOffset
* @param [out] dst output LocalTensor
* @param [in] src input LocalTensor
* @param [in] srcOffset input LocalTensor
* @param [in] srcBaseOffset base address of src
* @param [in] count element count
*/
template <typename T>
__aicore__ inline void Gather(const LocalTensor<T>& dst, const LocalTensor<T>& src,
const LocalTensor<uint32_t>& srcOffset, const uint32_t srcBaseOffset, const uint32_t count)
{
using PrimType = PrimT<T>;
#if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3002) || (__NPU_ARCH__ == 3102) || \
(__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
ASCENDC_ASSERT((SupportType<PrimType, uint8_t, int8_t, half, bfloat16_t, uint16_t, int16_t, float, uint32_t, int32_t>()),
{KERNEL_LOG(KERNEL_ERROR, "Failed to check dtype in Gather, current api support dtype combination is src and "
"dst both: uint8 / int8 / half / bfloat16_t / uint16_t / int16_t / float / uint32_t / int32_t");});
GatherImpl((__ubuf__ PrimType *)dst.GetPhyAddr(), (__ubuf__ PrimType *)src.GetPhyAddr(),
(__ubuf__ uint32_t *)srcOffset.GetPhyAddr(), srcBaseOffset, count);
#else
#if ASCENDC_CPU_DEBUG
if (!CheckFuncGather(dst, src, srcOffset, srcBaseOffset, count, "Gather")) {
ASCENDC_REPORT_CHECK_ERROR("Gather", KernelFuncType::CALCOUNT_MODE);
}
#endif
uint32_t vectorRegWidth = 256;
#if (__NPU_ARCH__ == 3003) || \
((__NPU_ARCH__ == 3113))
vectorRegWidth = VECTOR_REG_WIDTH;
#endif
uint32_t elementCountSingleRepeat;
if constexpr (sizeof(PrimType) == sizeof(uint16_t)) {
elementCountSingleRepeat = 128;
} else {
elementCountSingleRepeat = 64;
}
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3003) || \
(__NPU_ARCH__ == 3113)
elementCountSingleRepeat = vectorRegWidth / sizeof(T);
uint32_t repeatStride = vectorRegWidth / ONE_BLK_SIZE;
#endif
const uint32_t elementCountTail = count % elementCountSingleRepeat;
const uint8_t repeatTime = count / elementCountSingleRepeat;
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3003) || \
(__NPU_ARCH__ == 3113)
if (repeatTime > 0) {
Gather(dst, src, srcOffset, srcBaseOffset, (uint64_t)elementCountSingleRepeat, repeatTime,
repeatStride);
}
if (elementCountTail > 0) {
const uint32_t offset = count - elementCountTail;
Gather(dst[offset], src, srcOffset[offset], srcBaseOffset, (uint64_t)elementCountTail, 1,
repeatStride);
}
#else
if (repeatTime > 0) {
Gather(dst, src, srcOffset, srcBaseOffset, static_cast<uint64_t>(elementCountSingleRepeat), repeatTime,
DEFAULT_REPEAT_STRIDE);
}
if (elementCountTail > 0) {
const uint32_t offset = count - elementCountTail;
Gather(dst[offset], src, srcOffset[offset], srcBaseOffset, static_cast<uint64_t>(elementCountTail), 1,
DEFAULT_REPEAT_STRIDE);
}
#endif
#endif
}
}
#pragma end_pipe
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_VEC_GATHER_INTF_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_VEC_GATHER_INTF_IMPL_H__
#endif
