sqlalchemy/test/ext/mypy/plugin_files/sa_module_prefix.py-代码预览-sqlalchemy:基于 Python 的 SQL 工具包和对象关系映射项目

/**
 * 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.
 */
#include "aclnn_std.h"
#include <bitset>
#include "math/reduce_std_with_mean/op_api/reduce_std_with_mean.h"
#include "math/reduce_mean/op_api/reduce_mean.h"
#include "reduce_std_v2.h"
#include "aclnn_kernels/cast.h"
#include "aclnn_kernels/reshape.h"
#include "aclnn_kernels/contiguous.h"
#include "math/expand/op_api/expand.h"
#include "aclnn_kernels/transdata.h"
#include "opdev/common_types.h"
#include "opdev/data_type_utils.h"
#include "opdev/format_utils.h"
#include "opdev/op_dfx.h"
#include "opdev/op_executor.h"
#include "opdev/op_log.h"
#include "opdev/shape_utils.h"
#include "opdev/tensor_view_utils.h"
#include "aclnn_kernels/common/op_error_check.h"
#include "op_api/op_api_def.h"
#include "op_api/aclnn_check.h"
#include "opdev/platform.h"
#include "op_api/level2_base_caculation.h"

using namespace op;
using std::bitset;
#ifdef __cplusplus
extern "C" {
#endif

/* Std 算子的完整计算流程如下:
 *                   self
 *                    |
 *          Contiguous(workspace_0)
 *                    |
 *           ReduceMean(workspace_1)
 *                    |
 *       ReduceStdWithMean(workspace_2)
 *                    |
 *           Expand(workspace_3)
 *                    |
 *             Cast(workspace_4)
 *                    |
 *                ViewCopy
 *                    |
 *                 result
 */

constexpr size_t MAX_MASK_LEN = 64;
// 根据API定义，需要列出所能支持的所有dtype
static const std::initializer_list<op::DataType> ASCEND910_DTYPE_DTYPE_SUPPORT_LIST = {op::DataType::DT_FLOAT,
                                                                                       op::DataType::DT_FLOAT16};

static const std::initializer_list<op::DataType> ASCEND910B_DTYPE_DTYPE_SUPPORT_LIST = {
    op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_BF16};

static bool CheckDtypeValid(const aclTensor* self, const aclTensor* out)
{
    // 检查self的数据类型是否在std算子的支持列表内
    auto supportList = GetDtypeSupportListV2(ASCEND910B_DTYPE_DTYPE_SUPPORT_LIST, ASCEND910_DTYPE_DTYPE_SUPPORT_LIST);
    OP_CHECK_DTYPE_NOT_SUPPORT(self, supportList, return false);

    // 检查out的数据类型是否在std算子的支持列表内
    OP_CHECK_DTYPE_NOT_SUPPORT(out, supportList, return false);

    // 检查self的数据类型能否转换为输出的数据类型
    OP_CHECK_RESULT_DTYPE_CAST_FAILED(self->GetDataType(), out->GetDataType(), return false);

    return true;
}

static bool CheckDimValid(const aclTensor* self, const aclIntArray* dim)
{
    auto selfViewShape = self->GetViewShape();
    auto selfDimNum = static_cast<int64_t>(selfViewShape.GetDimNum());
    // 0维tensor
    if (selfDimNum == 0) {
        selfDimNum = 1;
    }
    uint64_t dimMask[64] = {0};
    // dim可以为空指针
    if (dim == nullptr) {
        return true;
    }
    // 获取dim元素
    for (size_t i = 0; i < dim->Size(); i++) {
        // dim值不能超出范围
        if (dim->operator[](i) >= selfDimNum || dim->operator[](i) < (-selfDimNum)) {
            OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Provided dim %ld must be in the range of [%ld, %ld].", dim->operator[](i),
                    -selfDimNum, selfDimNum - 1);
            return false;
        }
        // dim值可以为负
        if (dim->operator[](i) < 0) {
            if (dimMask[selfDimNum + dim->operator[](i)] == 1) {
                OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Dim %ld appears multiple times in the list of dims.",
                        selfDimNum + dim->operator[](i));
                return false;
            } else {
                dimMask[selfDimNum + dim->operator[](i)] = 1;
            }
            continue;
        }
        // dim值不能重复
        if (dimMask[dim->operator[](i)] == 1) {
            OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Dim %ld appears multiple times in the list of dims.", dim->operator[](i));
            return false;
        } else {
            dimMask[dim->operator[](i)] = 1;
        }
    }
    return true;
}

static void StdInferShape(const op::Shape& selfShape, const aclIntArray* dim, bool keepDim, op::Shape& reduceShape)
{
    bitset<MAX_MASK_LEN> dimMask = bitset<MAX_MASK_LEN>();
    for (size_t i = 0; i < dim->Size(); i++) {
        int64_t index = GetPosDimWithStd(dim->operator[](i), selfShape.GetDimNum());
        // 前序已检查， 此处如果dim不会重复
        dimMask.set(index);
    }

    for (size_t i = 0; i < selfShape.GetDimNum(); i++) {
        if (!dimMask[i]) {
            reduceShape.AppendDim(selfShape.GetDim(i));
        } else if (keepDim) {
            reduceShape.AppendDim(1);
        }
    }
}
static bool CheckShape(const aclTensor* self, const aclIntArray* dim, const bool keepDim, const aclTensor* out)
{
    OP_CHECK_MAX_DIM(self, MAX_SUPPORT_DIMS_NUMS, return false);
    OP_CHECK_MAX_DIM(out, MAX_SUPPORT_DIMS_NUMS, return false);
    // dim可以为空指针
    if (dim == nullptr || dim->Size() == 0) {
        return true;
    }
    op::Shape reduceShape;
    StdInferShape(self->GetViewShape(), dim, keepDim, reduceShape);

    // out的shape必须满足Infer shape
    OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE(out, reduceShape, return false);
    return true;
}

static aclnnStatus CheckParams(const aclTensor* self, const aclIntArray* dim, const bool keepdim, aclTensor* out)
{
    // 1. 检查参数是否为空指针
    CHECK_RET(CheckNotNull2Tensor(self, out), ACLNN_ERR_PARAM_NULLPTR);

    // 2. 检查输入的数据类型是否在API支持的数据类型范围之内，需要根据api定义校验
    CHECK_RET(CheckDtypeValid(self, out), ACLNN_ERR_PARAM_INVALID);

    // 3. 检查reduce的轴是否合理
    CHECK_RET(CheckDimValid(self, dim), ACLNN_ERR_PARAM_INVALID);

    // 4. 检查shape是否满足推导
    CHECK_RET(CheckShape(self, dim, keepdim, out), ACLNN_ERR_PARAM_INVALID);

    return ACLNN_SUCCESS;
}

static aclIntArray* CalcDim(const aclTensor* self, aclOpExecutor* executor)
{
    FVector<int64_t> dimVector;
    auto selfViewShape = self->GetViewShape();
    size_t selfDimNum = selfViewShape.GetDimNum();
    for (size_t i = 0; i < selfDimNum; i++) {
        dimVector.push_back(static_cast<int64_t>(i));
    }
    return executor->AllocIntArray(dimVector.data(), dimVector.size());
}

static aclIntArray* ConvToNotNegDim(const aclTensor* self, const aclIntArray* dim, aclOpExecutor* executor)
{
    FVector<int64_t> dimVector;
    auto selfViewShape = self->GetViewShape();
    int64_t selfDimNum = static_cast<int64_t>(selfViewShape.GetDimNum());
    int64_t dimValue;
    for (size_t i = 0; i < dim->Size(); i++) {
        dimValue = static_cast<int64_t>(dim->operator[](i));
        if (dimValue < 0) {
            dimVector.push_back(dimValue + selfDimNum);
        } else {
            dimVector.push_back(dimValue);
        }
    }
    return executor->AllocIntArray(dimVector.data(), dimVector.size());
}

static aclnnStatus aclnnStdV2ImplUnify(const aclTensor* self, const aclIntArray* dim, int64_t correction, bool keepdim,
                                       aclTensor* out, uint64_t* workspaceSize, UniqueExecutor& uniqueExecutor,
                                       aclOpExecutor** executor)
{
    bool isMeanOut = false;
    auto reduceStdV2Out = l0op::ReduceStdV2(self, dim, correction, keepdim, isMeanOut, uniqueExecutor.get());

    auto stdOut = std::get<0>(reduceStdV2Out);
    CHECK_RET(stdOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
    auto castOut = l0op::Cast(stdOut, out->GetDataType(), uniqueExecutor.get());
    CHECK_RET(castOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
    auto viewCopyResult = l0op::ViewCopy(castOut, out, uniqueExecutor.get());
    CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 获取计算过程中需要使用的workspace大小
    *workspaceSize = uniqueExecutor->GetWorkspaceSize();
    uniqueExecutor.ReleaseTo(executor);

    return ACLNN_SUCCESS;
}

aclnnStatus aclnnStdGetWorkspaceSize(const aclTensor* self, const aclIntArray* dim, const int64_t correction,
                                     bool keepdim, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor)
{
    OP_CHECK_COMM_INPUT(workspaceSize, executor);

    L2_DFX_PHASE_1(aclnnStd, DFX_IN(self, dim, correction, keepdim), DFX_OUT(out));

    // 固定写法，创建OpExecutor
    auto uniqueExecutor = CREATE_EXECUTOR();
    CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);

    // 固定写法，参数检查
    auto selfReshape = self;
    auto ret = CheckParams(selfReshape, dim, keepdim, out);
    CHECK_RET(ret == ACLNN_SUCCESS, ret);

    if (self->GetViewShape().GetDimNum() == 0) {
        int64_t selfShapeValue[1] = {1};
        aclIntArray* selfShape = uniqueExecutor.get()->AllocIntArray(selfShapeValue, 1);
        CHECK_RET(selfShape != nullptr, ACLNN_ERR_INNER_NULLPTR);
        auto selfReshapeI = l0op::Reshape(self, selfShape, uniqueExecutor.get());
        CHECK_RET(selfReshapeI != nullptr, ACLNN_ERR_INNER_NULLPTR);
        selfReshape = selfReshapeI;
    }

    const aclIntArray* dimArray;
    if (dim == nullptr || dim->Size() == 0) {
        // dim为空指针
        dimArray = CalcDim(self, uniqueExecutor.get());
    } else {
        // 负dim转成正值
        dimArray = ConvToNotNegDim(selfReshape, dim, uniqueExecutor.get());
    }
    CHECK_RET(dimArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
    // Std算子的空tensor在kernel中支持
    if (selfReshape->IsEmpty()) {
        // 空tensor填充NAN
        ret = CheckFillScalarShapeStdAndVar(out, NAN, uniqueExecutor.get());
        CHECK_RET(ret == ACLNN_SUCCESS, ret);
        *workspaceSize = uniqueExecutor->GetWorkspaceSize();
        uniqueExecutor.ReleaseTo(executor);
        return ret;
    }

    // shapeProd小于等于correction场景
    int64_t shapeProd = 1;
    shapeProd = CalcShapeProdStdAndVarMean(selfReshape, dimArray);
    if ((shapeProd == 1) && (shapeProd <= correction)) {
        // 返回NAN
        ret = CheckFillScalarShapeStdAndVar(out, NAN, uniqueExecutor.get());
        CHECK_RET(ret == ACLNN_SUCCESS, ret);
        *workspaceSize = uniqueExecutor->GetWorkspaceSize();
        uniqueExecutor.ReleaseTo(executor);
        return ret;
    }
    if ((correction > 1) && (shapeProd <= correction)) {
        // 返回INF
        ret = CheckFillScalarShapeStdAndVar(out, INFINITY, uniqueExecutor.get());
        CHECK_RET(ret == ACLNN_SUCCESS, ret);
        *workspaceSize = uniqueExecutor->GetWorkspaceSize();
        uniqueExecutor.ReleaseTo(executor);
        return ret;
    }

    // 固定写法，将输入self转换成连续的tensor
    auto selfContiguous = l0op::Contiguous(selfReshape, uniqueExecutor.get());
    CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);

    auto selfReformat = l0op::ReFormat(selfContiguous, Format::FORMAT_ND);
    CHECK_RET(selfReformat != nullptr, ACLNN_ERR_INNER_NULLPTR);

    if (IsRegBase()) {
        return aclnnStdV2ImplUnify(selfReformat, dimArray, correction, keepdim, out, workspaceSize, uniqueExecutor,
                                   executor);
    }

    // 调用Mean算子kernel
    auto meanOpOut = l0op::ReduceMean(selfContiguous, dimArray, true, uniqueExecutor.get());
    CHECK_RET(meanOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 调用Expand算子kernel
    FVector<int64_t> shapeVector;
    auto selfShape = selfReshape->GetViewShape();
    size_t selfDimNum = selfShape.GetDimNum();
    for (size_t i = 0; i < selfDimNum; i++) {
        shapeVector.emplace_back(selfShape[i]);
    }
    auto shapeArray = uniqueExecutor.get()->AllocIntArray(shapeVector.data(), selfDimNum);
    CHECK_RET(shapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
    auto expandOpOut = l0op::Expand(meanOpOut, shapeArray, uniqueExecutor.get());
    CHECK_RET(expandOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 调用StdWithMean算子kernel
    const aclTensor* stdWithMeanOpOut = nullptr;
    bool invert = false;
    float eps = 0.001f;
    stdWithMeanOpOut = l0op::ReduceStdWithMean(selfContiguous, expandOpOut, dimArray, correction, keepdim, invert, eps,
                                               uniqueExecutor.get());
    CHECK_RET(stdWithMeanOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 固定写法，将计算结果转换成输出out的数据类型
    auto castOut = l0op::Cast(stdWithMeanOpOut, out->GetDataType(), uniqueExecutor.get());
    CHECK_RET(castOut != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 固定写法，将计算结果拷贝到输出out上，out可能是非连续的tensor
    auto viewCopyResult = l0op::ViewCopy(castOut, out, uniqueExecutor.get());
    CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);

    // 固定写法，获取计算过程中需要使用的workspace大小
    *workspaceSize = uniqueExecutor->GetWorkspaceSize();
    uniqueExecutor.ReleaseTo(executor); // 需要把 uniqueExecutor持有executor转移给executor
    return ACLNN_SUCCESS;
}

aclnnStatus aclnnStd(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, const aclrtStream stream)
{
    L2_DFX_PHASE_2(aclnnStd);
    // 固定写法，调用框架能力，完成计算
    return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
}

#ifdef __cplusplus
}
#endif