@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.5

package/assets/ops-direct-invoke/skills/ascendc-performance-best-practices/references/reduce/templates/dav310/softmax_v2_ar_full_load.template

@@ -0,0 +1,429 @@
+/**
+ * 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 softmax_v2_ar_full_load.h
+ * \brief
+ */
+
+#ifndef SOFTMAX_V2_AR_FULL_LOAD_H
+#define SOFTMAX_V2_AR_FULL_LOAD_H
+
+#include "kernel_tiling/kernel_tiling.h"
+#include "kernel_operator.h"
+
+#include "softmax_v2_base.h"
+
+#ifndef INFINITY
+#define INFINITY (__builtin_inff())
+#endif
+
+namespace SoftmaxV2Ops
+{
+using namespace AscendC;
+
+constexpr static uint32_t DOUBLE_BUFFER = 2;
+constexpr static uint32_t BLOCK_SIZE = 32;  // 32B
+constexpr static uint32_t AR_FULL_LOAD_BINARY_TMP_BYTES = 512;
+
+template <typename T_in, typename T_out>
+class SoftmaxV2AR : public SoftmaxV2OpsBase
+{
+public:
+    __aicore__ inline SoftmaxV2AR(TPipe* pipe)
+    {
+        pipe_ = pipe;
+    };
+
+    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, const SoftmaxV2ARTilingData* tilingData);
+    __aicore__ inline void Process();
+
+private:
+    __aicore__ inline void ProcessUB(int64_t ubA, int64_t aOffset);
+
+    __aicore__ inline void FirstNormCompute(int64_t ubA, __local_mem__ T_in* xInAddr,
+                                            __local_mem__ float* xTmpLocalAddr);
+    __aicore__ inline void SecondNormCompute(const LocalTensor<T_out>& dstTensor, const LocalTensor<float>& srcTensor,
+                                             const LocalTensor<float>& reduceSumTempTensor, const int64_t aSize,
+                                             const int64_t rSize, const int64_t stride);
+    __aicore__ inline void SecondNormComputePost(const LocalTensor<T_out>& dstTensor,
+                                                 const LocalTensor<float>& srcTensor,
+                                                 const LocalTensor<float>& oriSrcTensor, const int64_t aSize,
+                                                 const int64_t rSize, const int64_t stride);
+
+    __aicore__ inline void CopyInX(const LocalTensor<T_in>& xInUb, int64_t ubA, int64_t offset);
+    __aicore__ inline void CopyOutY(const LocalTensor<T_out>& yOutUb, int64_t ubA, int64_t offset);
+    __aicore__ inline void LoadTensorForDtypeTIn(__local_mem__ T_in* src, AscendC::MicroAPI::RegTensor<float>& dst,
+                                                 AscendC::MicroAPI::MaskReg& preg, uint32_t offset);
+    __aicore__ inline void StoreTensorForDtypeTOut(__local_mem__ T_out* dst, AscendC::MicroAPI::RegTensor<float>& src,
+                                                   AscendC::MicroAPI::MaskReg& preg, uint32_t offset);
+
+private:
+    /* global memory address */
+    GlobalTensor<T_in> xGm_;
+    GlobalTensor<T_out> yGm_;
+
+    /* ascendc variable */
+    TPipe* pipe_ = nullptr;
+    TQue<QuePosition::VECIN, 1> xQueue_;
+    TQue<QuePosition::VECOUT, 1> yQueue_;
+
+    TBuf<> tmpLocalBuffer_;
+
+    int64_t blockA_ = 0;  // 获取分块操作中的单个块的大小
+    const SoftmaxV2ARTilingData* tl_ = nullptr;
+};
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::Init(GM_ADDR x, GM_ADDR y, const SoftmaxV2ARTilingData* tilingData)
+{
+    this->tl_ = tilingData;
+    // GM not need align.
+    // 获取分块操作中的单个块的大小。判断是否是最后一块，是最后一块，则等于剩余元素的数量，否则等于固定的单核处理的行数
+    this->blockA_ = (AscendC::GetBlockIdx() == AscendC::GetBlockNum() - 1)
+                        ? (tl_->a - tl_->aBlockFactor * (AscendC::GetBlockNum() - 1))
+                        : tl_->aBlockFactor;
+    int64_t aGmOffset = tl_->aBlockFactor * AscendC::GetBlockIdx() * tl_->r;
+    // 初始化GM Tensor
+    xGm_.SetGlobalBuffer((__gm__ T_in*)x + aGmOffset);
+    yGm_.SetGlobalBuffer((__gm__ T_out*)y + aGmOffset);
+
+    // 初始化Pipe
+    int64_t ubBufferSize = tl_->ubFactor * tl_->rAligned;
+    pipe_->InitBuffer(this->xQueue_, DOUBLE_BUFFER, ubBufferSize * sizeof(T_in));
+    pipe_->InitBuffer(this->yQueue_, DOUBLE_BUFFER, ubBufferSize * sizeof(T_out));
+    pipe_->InitBuffer(this->tmpLocalBuffer_, ubBufferSize * sizeof(float) + AR_FULL_LOAD_BINARY_TMP_BYTES);
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::Process()
+{
+    // ubLoop: 表示需要多少个子块来覆盖singleA大小的数据
+    int64_t ubLoop = ops::CeilDiv(this->blockA_, tl_->ubFactor);
+    int64_t lastUbFactor = this->blockA_ - tl_->ubFactor * (ubLoop - 1);
+    // 循环处理每个子块
+    for (int64_t ubLoopIdx = 0; ubLoopIdx < ubLoop; ubLoopIdx++) {
+        // aOffset：计算当前子块的偏移量
+        int64_t aOffset = ubLoopIdx * tl_->ubFactor * tl_->r;
+        int64_t ubA = (ubLoopIdx == (ubLoop - 1)) ? lastUbFactor : tl_->ubFactor;
+        ProcessUB(ubA, aOffset);
+    }
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::ProcessUB(int64_t ubA, int64_t aOffset)
+{
+    LocalTensor<T_in> xInUb = xQueue_.AllocTensor<T_in>();
+    CopyInX(xInUb, ubA, aOffset);
+    xQueue_.EnQue(xInUb);
+    xInUb = xQueue_.DeQue<T_in>();
+
+    LocalTensor<float> tmpLocalTensor = tmpLocalBuffer_.AllocTensor<float>();
+    LocalTensor<float> binaryTmpLocalTensor = tmpLocalTensor[tl_->ubFactor * tl_->rAligned];
+    LocalTensor<float> xTmpLocalTensor = tmpLocalTensor[0];
+
+    __local_mem__ T_in* xInUbAddr = (__local_mem__ T_in*)xInUb.GetPhyAddr();
+    __local_mem__ float* xTmpLocalAddr = (__local_mem__ float*)xTmpLocalTensor.GetPhyAddr();
+    __local_mem__ float* binaryTmpLocalAddr = (__local_mem__ float*)binaryTmpLocalTensor.GetPhyAddr();
+
+    FirstNormCompute(ubA, xInUbAddr, xTmpLocalAddr);
+    xQueue_.FreeTensor<T_in>(xInUb);
+    LocalTensor<T_out> yInUb = yQueue_.AllocTensor<T_out>();
+    SecondNormCompute(yInUb, xTmpLocalTensor, binaryTmpLocalTensor, ubA, tl_->r, tl_->rAligned);
+    yQueue_.EnQue(yInUb);
+
+    LocalTensor<T_out> yOutUb = yQueue_.DeQue<T_out>();
+    CopyOutY(yOutUb, ubA, aOffset);
+    yQueue_.FreeTensor(yOutUb);
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::FirstNormCompute(int64_t ubA, __local_mem__ T_in* xInAddr,
+                                                                  __local_mem__ float* xTmpLocalAddr)
+{
+    int64_t rAligned = tl_->rAligned;
+    int64_t ubActualR = tl_->r;
+    uint16_t ubActualA = static_cast<uint16_t>(ubA);
+    int64_t tailUbBlockSize = tl_->r - VL_FP32 * (tl_->rLoopCount - 1);
+    uint16_t rLoopCount = static_cast<uint16_t>(tl_->rLoopCount);
+    uint16_t rLoopCountTmp = rLoopCount - 1;
+
+    __VEC_SCOPE__
+    {
+        AscendC::MicroAPI::RegTensor<float> vreg0;
+        AscendC::MicroAPI::RegTensor<float> vreg1;
+        AscendC::MicroAPI::RegTensor<float> vreg2;
+        AscendC::MicroAPI::RegTensor<float> vreg3;
+        AscendC::MicroAPI::RegTensor<float> vreg4;
+        AscendC::MicroAPI::RegTensor<float> vreg5;
+
+        AscendC::MicroAPI::MaskReg maskTail;
+        AscendC::MicroAPI::MaskReg maskSub;
+        AscendC::MicroAPI::MaskReg maskMax;
+        AscendC::MicroAPI::MaskReg maskAll;
+
+        uint32_t sreg0 = tailUbBlockSize;
+        uint32_t sreg1 = VL_FP32;
+        maskTail = AscendC::MicroAPI::UpdateMask<uint32_t>(sreg0);
+        maskAll = AscendC::MicroAPI::UpdateMask<uint32_t>(sreg1);
+
+        for (uint16_t k = 0; k < ubActualA; k++) {
+            uint32_t tailAddrPtr = k * rAligned + VL_FP32 * (rLoopCount - 1);
+            AscendC::MicroAPI::Duplicate(vreg0, static_cast<float>(-INFINITY), maskAll);
+            LoadTensorForDtypeTIn(xInAddr, vreg1, maskTail, tailAddrPtr);
+            AscendC::MicroAPI::Max(vreg1, vreg0, vreg1, maskTail);
+            AscendC::MicroAPI::Copy<float, AscendC::MicroAPI::MaskMergeMode::MERGING>(vreg0, vreg1, maskTail);
+
+            uint32_t sreg2 = static_cast<uint32_t>(ubActualR - tailUbBlockSize);
+            for (uint16_t i = 0; i < rLoopCountTmp; i++) {
+                maskMax = AscendC::MicroAPI::UpdateMask<float>(sreg2);
+                uint32_t addrPtr = i * VL_FP32 + k * rAligned;
+                LoadTensorForDtypeTIn(xInAddr, vreg1, maskMax, addrPtr);
+                AscendC::MicroAPI::Max(vreg0, vreg0, vreg1, maskMax);
+            }
+            AscendC::MicroAPI::ReduceMax(vreg2, vreg0, maskAll);
+            AscendC::MicroAPI::Duplicate(vreg3, vreg2, maskAll);
+
+            // 求sub和exp
+            uint32_t sreg3 = ubActualR;
+            for (uint16_t i = 0; i < rLoopCount; i++) {
+                maskSub = AscendC::MicroAPI::UpdateMask<float>(sreg3);
+                uint32_t addrPtr = i * VL_FP32 + k * rAligned;
+                LoadTensorForDtypeTIn(xInAddr, vreg1, maskSub, addrPtr);
+                AscendC::MicroAPI::Sub(vreg4, vreg1, vreg3, maskSub);
+                AscendC::MicroAPI::Exp(vreg5, vreg4, maskSub);
+                AscendC::MicroAPI::DataCopy(((__local_mem__ float*)xTmpLocalAddr + addrPtr), vreg5, maskSub);
+            }
+        }
+    }
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::SecondNormComputePost(const LocalTensor<T_out>& dstTensor,
+                                                                       const LocalTensor<float>& srcTensor,
+                                                                       const LocalTensor<float>& oriSrcTensor,
+                                                                       const int64_t aSize, const int64_t rSize,
+                                                                       const int64_t stride)
+{
+    if (aSize <= 0) {
+        return;
+    }
+    if (rSize <= 0) {
+        return;
+    }
+    if (rSize > CONST_TWO * VL_FP32) {
+        return;
+    }
+
+    uint16_t loopTimes = aSize;
+    uint16_t rLoopCount = tl_->rLoopCount;
+    uint16_t oriR = tl_->r;
+    uint16_t oriRAligned = tl_->rAligned;
+
+    if (rSize <= VL_FP32) {
+        __local_mem__ T_out* dst = (__local_mem__ T_out*)dstTensor.GetPhyAddr();
+        __local_mem__ float* src = (__local_mem__ float*)srcTensor.GetPhyAddr();
+        __local_mem__ float* oriSrc = (__local_mem__ float*)oriSrcTensor.GetPhyAddr();
+
+        __VEC_SCOPE__
+        {
+            uint32_t count = static_cast<uint32_t>(rSize);
+            AscendC::MicroAPI::RegTensor<float> aReg, bReg, cReg, dReg;
+            AscendC::MicroAPI::MaskReg pMask = AscendC::MicroAPI::UpdateMask<float>(count);
+            AscendC::MicroAPI::MaskReg pFull =
+                AscendC::MicroAPI::CreateMask<float, AscendC::MicroAPI::MaskPattern::ALL>();
+            AscendC::MicroAPI::MaskReg maskOri;
+            for (uint16_t i = 0; i < loopTimes; ++i) {
+                DataCopy(aReg, (__local_mem__ float*)src + i * static_cast<uint32_t>(stride));
+                ReduceSum(bReg, aReg, pMask);
+                Duplicate(cReg, bReg, pFull);
+                uint32_t sreg0 = static_cast<uint32_t>(oriR);
+                for (uint16_t j = 0; j < rLoopCount; ++j) {
+                    maskOri = AscendC::MicroAPI::UpdateMask<float>(sreg0);
+                    uint32_t addrPtr = j * VL_FP32 + i * oriRAligned;
+                    DataCopy(dReg, ((__local_mem__ float*)oriSrc + addrPtr));
+                    Div(dReg, dReg, cReg, maskOri);
+                    StoreTensorForDtypeTOut(dst, dReg, maskOri, addrPtr);
+                }
+            }
+        }
+    } else {
+        __local_mem__ T_out* dst = (__local_mem__ T_out*)dstTensor.GetPhyAddr();
+        __local_mem__ float* src0 = (__local_mem__ float*)srcTensor.GetPhyAddr();
+        __local_mem__ float* src1 = (__local_mem__ float*)srcTensor.GetPhyAddr() + VL_FP32;
+        __local_mem__ float* oriSrc = (__local_mem__ float*)oriSrcTensor.GetPhyAddr();
+
+        __VEC_SCOPE__
+        {
+            uint32_t count = static_cast<uint32_t>(rSize - VL_FP32);
+            AscendC::MicroAPI::RegTensor<float> aReg, bReg, cReg, dReg, eReg;
+            AscendC::MicroAPI::MaskReg pMask = AscendC::MicroAPI::UpdateMask<float>(count);
+            AscendC::MicroAPI::MaskReg pFull =
+                AscendC::MicroAPI::CreateMask<float, AscendC::MicroAPI::MaskPattern::ALL>();
+            AscendC::MicroAPI::MaskReg maskOri;
+            for (uint16_t i = 0; i < loopTimes; ++i) {
+                DataCopy(aReg, (__local_mem__ float*)src0 + i * static_cast<uint32_t>(stride));
+                DataCopy(bReg, (__local_mem__ float*)src1 + i * static_cast<uint32_t>(stride));
+                Add<float, AscendC::MicroAPI::MaskMergeMode::ZEROING>(cReg, aReg, bReg, pMask);
+                Copy<float, AscendC::MicroAPI::MaskMergeMode::MERGING>(aReg, cReg, pMask);
+                ReduceSum(bReg, aReg, pFull);
+                Duplicate(dReg, bReg, pFull);
+                uint32_t sreg0 = static_cast<uint32_t>(oriR);
+                for (uint16_t j = 0; j < rLoopCount; ++j) {
+                    maskOri = AscendC::MicroAPI::UpdateMask<float>(sreg0);
+                    uint32_t addrPtr = j * VL_FP32 + i * oriRAligned;
+                    DataCopy(eReg, ((__local_mem__ float*)oriSrc + addrPtr));
+                    Div(eReg, eReg, dReg, maskOri);
+                    StoreTensorForDtypeTOut(dst, eReg, maskOri, addrPtr);
+                }
+            }
+        }
+    }
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::SecondNormCompute(const LocalTensor<T_out>& dstTensor,
+                                                                   const LocalTensor<float>& srcTensor,
+                                                                   const LocalTensor<float>& reduceSumTempTensor,
+                                                                   const int64_t aSize, const int64_t rSize,
+                                                                   const int64_t stride)
+{
+    if (aSize <= 0) {
+        return;
+    }
+    if (rSize <= 0) {
+        return;
+    }
+    if (rSize <= CONST_TWO * VL_FP32) {
+        SecondNormComputePost(dstTensor, srcTensor, srcTensor, aSize, rSize, stride);
+        return;
+    }
+
+    int64_t ceilVLCount =
+        ops::CeilDiv(static_cast<int64_t>(rSize * sizeof(float)), static_cast<int64_t>(GetVRegSize()));
+    int64_t floorVLCount =
+        ops::FloorDiv(static_cast<int64_t>(rSize * sizeof(float)), static_cast<int64_t>(GetVRegSize()));
+    int64_t foldPoint = FindNearestPower2(ceilVLCount);
+
+    uint16_t outerLoopTimes = aSize;
+    uint16_t tailFoldLoopTimes = ceilVLCount - floorVLCount;
+    uint32_t tailFoldElemCount = static_cast<uint32_t>(rSize - floorVLCount * VL_FP32);
+    uint16_t mainFoldLoopTimes = floorVLCount - foldPoint;
+    uint16_t unFoldLoopTimes = foldPoint + foldPoint - ceilVLCount;
+    uint32_t outerLoopStride = stride;
+    uint32_t innerLoopStride = VL_FP32;
+    uint32_t outerLoopDstStride =
+        ops::Aligned(static_cast<int64_t>(foldPoint), static_cast<int64_t>(GetUbBlockSize() / sizeof(float)));
+
+    int64_t foldSrcBOffset = foldPoint * VL_FP32;
+    int64_t tailSrcAOffset = mainFoldLoopTimes * VL_FP32;
+    int64_t tailSrcBOffset = floorVLCount * VL_FP32;
+    int64_t unFoldSrcOffset = (mainFoldLoopTimes + tailFoldLoopTimes) * VL_FP32;
+
+    __local_mem__ float* dst = (__local_mem__ float*)reduceSumTempTensor.GetPhyAddr();
+    __local_mem__ float* foldSrcA = (__local_mem__ float*)srcTensor.GetPhyAddr();
+    __local_mem__ float* foldSrcB = (__local_mem__ float*)srcTensor.GetPhyAddr() + foldSrcBOffset;
+    __local_mem__ float* tailSrcA = (__local_mem__ float*)srcTensor.GetPhyAddr() + tailSrcAOffset;
+    __local_mem__ float* tailSrcB = (__local_mem__ float*)srcTensor.GetPhyAddr() + tailSrcBOffset;
+    __local_mem__ float* unFoldSrc = (__local_mem__ float*)srcTensor.GetPhyAddr() + unFoldSrcOffset;
+
+    __VEC_SCOPE__
+    {
+        AscendC::MicroAPI::MaskReg pFull = AscendC::MicroAPI::CreateMask<float, AscendC::MicroAPI::MaskPattern::ALL>();
+        AscendC::MicroAPI::UnalignReg UReg;
+
+        for (uint16_t i = 0; i < outerLoopTimes; ++i) {
+            dst = (__local_mem__ float*)reduceSumTempTensor.GetPhyAddr() + i * outerLoopDstStride;
+            for (uint16_t j = 0; j < mainFoldLoopTimes; ++j) {
+                AscendC::MicroAPI::RegTensor<float> aReg, bReg, cReg, dReg;
+                DataCopy(aReg, (__local_mem__ float*)foldSrcA + i * outerLoopStride + j * innerLoopStride);
+                DataCopy(bReg, (__local_mem__ float*)foldSrcB + i * outerLoopStride + j * innerLoopStride);
+                Add<float, AscendC::MicroAPI::MaskMergeMode::ZEROING>(cReg, aReg, bReg, pFull);
+                ReduceSum(dReg, cReg, pFull);
+                AscendC::MicroAPI::DataCopyUnAlign((__local_mem__ float*&)dst, dReg, UReg, 1);
+            }
+            for (uint16_t j = 0; j < tailFoldLoopTimes; ++j) {
+                uint32_t count = static_cast<uint32_t>(tailFoldElemCount);
+                AscendC::MicroAPI::RegTensor<float> aReg, bReg, cReg;
+                AscendC::MicroAPI::MaskReg pMask = AscendC::MicroAPI::UpdateMask<float>(count);
+                DataCopy(aReg, (__local_mem__ float*)tailSrcA + i * outerLoopStride + j * innerLoopStride);
+                DataCopy(bReg, (__local_mem__ float*)tailSrcB + i * outerLoopStride + j * innerLoopStride);
+                Add<float, AscendC::MicroAPI::MaskMergeMode::ZEROING>(cReg, aReg, bReg, pMask);
+                Copy<float, AscendC::MicroAPI::MaskMergeMode::MERGING>(aReg, cReg, pMask);
+                ReduceSum(bReg, aReg, pFull);
+                AscendC::MicroAPI::DataCopyUnAlign((__local_mem__ float*&)dst, bReg, UReg, 1);
+            }
+            for (uint16_t j = 0; j < unFoldLoopTimes; ++j) {
+                AscendC::MicroAPI::RegTensor<float> aReg, bReg;
+                DataCopy(aReg, (__local_mem__ float*)unFoldSrc + i * outerLoopStride + j * innerLoopStride);
+                ReduceSum(bReg, aReg, pFull);
+                AscendC::MicroAPI::DataCopyUnAlign((__local_mem__ float*&)dst, bReg, UReg, 1);
+            }
+            AscendC::MicroAPI::DataCopyUnAlignPost((__local_mem__ float*&)dst, UReg, 0);
+        }
+    }
+    SecondNormComputePost(dstTensor, reduceSumTempTensor, srcTensor, aSize, foldPoint, outerLoopDstStride);
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::LoadTensorForDtypeTIn(__local_mem__ T_in* src,
+                                                                       AscendC::MicroAPI::RegTensor<float>& dst,
+                                                                       AscendC::MicroAPI::MaskReg& preg,
+                                                                       uint32_t offset)
+{
+    if constexpr (IsSameType<T_in, float>::value) {
+        DataCopy<float, AscendC::MicroAPI::LoadDist::DIST_NORM>(dst, src + offset);
+    } else {
+        AscendC::MicroAPI::RegTensor<T_in> xFp16;
+        DataCopy<T_in, AscendC::MicroAPI::LoadDist::DIST_UNPACK_B16>(xFp16, src + offset);
+        Cast<float, T_in, castTraitFp16ToFp32>(dst, xFp16, preg);
+    }
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::StoreTensorForDtypeTOut(__local_mem__ T_out* dst,
+                                                                         AscendC::MicroAPI::RegTensor<float>& src,
+                                                                         AscendC::MicroAPI::MaskReg& preg,
+                                                                         uint32_t offset)
+{
+    if constexpr (IsSameType<T_out, float>::value) {
+        DataCopy<T_out, AscendC::MicroAPI::StoreDist::DIST_NORM>(dst + offset, src, preg);
+    } else {
+        AscendC::MicroAPI::RegTensor<T_out> xFp16;
+        Cast<T_out, float, castTraitFp32ToFp16>(xFp16, src, preg);
+        DataCopy<T_out, AscendC::MicroAPI::StoreDist::DIST_PACK_B32>(dst + offset, xFp16, preg);
+    }
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::CopyInX(const LocalTensor<T_in>& xInUb, int64_t ubA, int64_t offset)
+{
+    DataCopyPadParams padParams{false, 0, 0, 0};
+    DataCopyParams copyInParams;
+    copyInParams.blockCount = ubA;
+    copyInParams.blockLen = tl_->r * sizeof(T_in);
+    copyInParams.srcStride = 0;
+    copyInParams.dstStride = (tl_->rAligned - tl_->r) * sizeof(T_in) / BLOCK_SIZE;
+    DataCopyPad(xInUb, xGm_[offset], copyInParams, padParams);
+}
+
+template <typename T_in, typename T_out>
+__aicore__ inline void SoftmaxV2AR<T_in, T_out>::CopyOutY(const LocalTensor<T_out>& yOutUb, int64_t ubA, int64_t offset)
+{
+    DataCopyParams copyOutParams;
+    copyOutParams.blockCount = ubA;
+    copyOutParams.blockLen = tl_->r * sizeof(T_out);
+    copyOutParams.srcStride = (tl_->rAligned - tl_->r) * sizeof(T_out) / BLOCK_SIZE;
+    copyOutParams.dstStride = 0;
+    DataCopyPad(yGm_[offset], yOutUb, copyOutParams);
+}
+
+}  // namespace SoftmaxV2Ops
+#endif  // SOFTMAX_V2_AR_FULL_LOAD_H