#include <iostream>
#include <vector>
#include <getopt.h>
#include <unistd.h>
#include <dirent.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <ctime>
#include <sys/wait.h>
#include <dlfcn.h>
#include <cstdio>
#include <string>
#include <vector>
#include <iomanip>
#include <string.h>
#include <list>
#include <algorithm>
#include <stdarg.h>
#include "acl/acl.h"
#include "aclnn/aclnn_base.h"
#include "acl/acl_base.h"
#include "acl/acl_rt.h"
#include "sec_common.h"
#include "utils.h"

#include "../fuzzlib/secodeFuzz.h"
#include "aclnnop/aclnn_binary_cross_entropy_with_logits.h"

using namespace std;



/* ========== [CONSISTENT] 一致性压测辅助函数（gen_consistent_from_normal 生成） ========== */
static const int CONSISTENT_MAX_RUN_TIMES = 64; /* hash数组上限，防越界 */
static int CONSISTENT_RUN_TIMES = 2; /* 每用例执行次数：默认2次，对比2次输出是否一致；经环境变量 CONSISTENT_RUN_TIMES 覆盖 */

static const char* ConsistentDtypeStr(aclDataType t) {
    switch (t) {
        case ACL_FLOAT: return "FLOAT"; case ACL_FLOAT16: return "FLOAT16";
        case ACL_BF16: return "BFLOAT16"; case ACL_DOUBLE: return "DOUBLE";
        case ACL_INT8: return "INT8"; case ACL_INT16: return "INT16";
        case ACL_INT32: return "INT32"; case ACL_INT64: return "INT64";
        case ACL_UINT8: return "UINT8"; case ACL_UINT16: return "UINT16";
        case ACL_UINT32: return "UINT32"; case ACL_UINT64: return "UINT64";
        case ACL_BOOL: return "BOOL"; case ACL_COMPLEX64: return "COMPLEX64";
        case ACL_COMPLEX128: return "COMPLEX128";
        default: return "UNKNOWN";
    }
}

static const char* ConsistentFormatStr(aclFormat f) {
    switch (f) {
        case ACL_FORMAT_NCHW: return "NCHW"; case ACL_FORMAT_NHWC: return "NHWC";
        case ACL_FORMAT_ND: return "ND"; case ACL_FORMAT_NC1HWC0: return "NC1HWC0";
        case ACL_FORMAT_FRACTAL_Z: return "FRACTAL_Z"; case ACL_FORMAT_HWCN: return "HWCN";
        case ACL_FORMAT_NDHWC: return "NDHWC"; case ACL_FORMAT_FRACTAL_NZ: return "FRACTAL_NZ";
        case ACL_FORMAT_NCDHW: return "NCDHW"; case ACL_FORMAT_NC: return "NC";
        case ACL_FORMAT_NCL: return "NCL";
        default: return "OTHER";
    }
}

static size_t ConsistentSizeOfDtype(aclDataType dataType) {
    switch (dataType) {
        case ACL_FLOAT: return 4; case ACL_FLOAT16: return 2;
        case ACL_INT8: return 1; case ACL_INT32: return 4;
        case ACL_UINT8: return 1; case ACL_INT16: return 2;
        case ACL_UINT16: return 2; case ACL_UINT32: return 4;
        case ACL_INT64: return 8; case ACL_UINT64: return 8;
        case ACL_DOUBLE: return 8; case ACL_BOOL: return 1;
        case ACL_BF16: return 2; case ACL_COMPLEX64: return 8;
        case ACL_COMPLEX128: return 16;
        /* 以下 dtype 的 sizeof(T)=1（vector<int8_t>），必须显式列出，禁止落入 default:4 */
        case ACL_INT4: return 1;
        case ACL_HIFLOAT8: return 1;
        case ACL_FLOAT8_E5M2: return 1;
        case ACL_FLOAT8_E4M3FN: return 1;
        case ACL_FLOAT8_E8M0: return 1;
        case ACL_FLOAT4_E2M1: return 1;
        case ACL_FLOAT4_E1M2: return 1;
        default: return 0; /* 未知 dtype 返回 0，触发 [FUZZ_CONSISTENT_BYTES] 告警 */
    }
}

static void ConsistentPrintShape(const std::vector<int64_t>& shape) {
    printf("[");
    for (size_t k = 0; k < shape.size(); k++) {
        printf("%lld%s", (long long)shape[k], (k + 1 < shape.size()) ? "," : "");
    }
    printf("]");
}

static uint64_t ConsistentHashBuffer(const void* data, size_t len) {
    const uint8_t* p = (const uint8_t*)data;
    uint64_t hash = 1469598103934665603ULL; /* FNV-1a offset basis */
    for (size_t k = 0; k < len; k++) {
        hash ^= (uint64_t)p[k];
        hash *= 1099511628211ULL; /* FNV-1a prime */
    }
    return hash;
}

static uint64_t ConsistentHashDeviceTensor(void* deviceAddr, size_t bytes) {
    if (deviceAddr == nullptr || bytes == 0) return 0;
    void* hostBuf = malloc(bytes);
    if (hostBuf == nullptr) return 0;
    if (aclrtMemcpy(hostBuf, bytes, deviceAddr, bytes, ACL_MEMCPY_DEVICE_TO_HOST) != ACL_SUCCESS) {
        free(hostBuf);
        return 0;
    }
    uint64_t h = ConsistentHashBuffer(hostBuf, bytes);
    free(hostBuf);
    return h;
}
/* ========== [CONSISTENT] 辅助函数结束 ========== */

int aclnnBinaryCrossEntropyWithLogits_consistent(int fuzz_num)
{
    auto ret = Init(deviceId, &context, &stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // [CONSISTENT] md文档标注默认确定性实现，无需打开确定性计算开关
    printf("[FUZZ_DETERMINISTIC] api=aclnnBinaryCrossEntropyWithLogits, enable_deterministic=0, reason=md文档标注：aclnnBinaryCrossEntropyWithLogits 默认确定性实现。\n");
    WriteDTLog("aclnnBinaryCrossEntropyWithLogits", "aclnnBinaryCrossEntropyWithLogits consistent op fuzz start \n");

    int dtype0Enum[] = {0, 1, 27};
    int dtype0Num = sizeof(dtype0Enum) / sizeof(dtype0Enum[0]);

    aclrtSetOpExecuteTimeOut(1091);
    DT_Set_TimeOut_Second(3000);
    DT_FUZZ_START(0, fuzz_num, (char*)__func__, 0)
    {
        int i = 0;
        void* selfDeviceAddr = nullptr;
        void* targetDeviceAddr = nullptr;
        void* weightDeviceAddr = nullptr;
        void* posWeightDeviceAddr = nullptr;
        void* outDeviceAddr = nullptr;
        aclTensor* self = nullptr;
        aclTensor* target = nullptr;
        aclTensor* weight = nullptr;
        aclTensor* posWeight = nullptr;
        aclTensor* out = nullptr;
        uint64_t workspaceSize = 0;
        aclOpExecutor* executor = nullptr;
        void* workspaceAddr = nullptr;

        // [CONSISTENT] 一致性压测：固定输入重复执行 + hash 对比用变量
        char curTilingKey[64];
        snprintf(curTilingKey, sizeof(curTilingKey), "N/A");
        int cstRunCount = 0;
        bool cstAllOk = true;
        bool cstResultPrinted = false;  // 判定路径是否已打印 FUZZ_RESULT（去重 ERROR 标签处打印）
        uint64_t cstSelfShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstSelfDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstSelfBytes = 0;
        void* cstSelfSaveBuf = nullptr;
        uint64_t cstSelfHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstSelfHashArr, 0, sizeof(cstSelfHashArr));
        uint64_t cstTargetShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstTargetDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstTargetBytes = 0;
        void* cstTargetSaveBuf = nullptr;
        uint64_t cstTargetHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstTargetHashArr, 0, sizeof(cstTargetHashArr));
        uint64_t cstWeightShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstWeightDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstWeightBytes = 0;
        void* cstWeightSaveBuf = nullptr;
        uint64_t cstWeightHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstWeightHashArr, 0, sizeof(cstWeightHashArr));
        uint64_t cstPosWeightShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstPosWeightDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstPosWeightBytes = 0;
        void* cstPosWeightSaveBuf = nullptr;
        uint64_t cstPosWeightHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstPosWeightHashArr, 0, sizeof(cstPosWeightHashArr));
        uint64_t cstOutShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstOutDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstOutBytes = 0;
        void* cstOutSaveBuf = nullptr;
        uint64_t cstOutHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstOutHashArr, 0, sizeof(cstOutHashArr));

        printf("[FUZZ_LOG] plog_file: aclnnBinaryCrossEntropyWithLogits_consistent_plog_case_%d\n", fuzzi);
        LOG_PRINT("aclnnBinaryCrossEntropyWithLogits_consistent %d\n", fuzzi);
        try
        {

            // === Tensor Size distribution control ===

            int sizeBucket = *(int *)DTSetGetNumberRange(&g_Element[i++], 1, 0, 1, "i");  // [CONSISTENT] 收紧size档位(仅小/中规模), 避免大shape快照导致host OOM

            long long totalSizeMin = 0, totalSizeMax = 0;

            switch (sizeBucket) {

                case 0: totalSizeMin = 1;          totalSizeMax = 1000;          break;

                case 1: totalSizeMin = 1001;       totalSizeMax = 67108864LL;  break;  // [CONSISTENT] 中档上限收紧至64M元素(快照双份内存峰值约3GB)

                case 2: totalSizeMin = 1073741825LL; totalSizeMax = 2147483648LL; break;

                case 3: totalSizeMin = 2147483649LL; totalSizeMax = 4294967296LL; break;

            }

            int shape1 = *(int *)DTSetGetNumberRange(&g_Element[i++], 4, 1, MAX_SHAPE_SIZE, "i");
            int shape2 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape3 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape4 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape5 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape6 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape7 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            int shape8 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 1, MAX_SHAPE_SIZE, "i");
            auto dtype0 = (aclDataType)*(int *)DT_SetGetNumberEnumV3(i++, aclDataType::ACL_FLOAT, dtype0Enum, dtype0Num);
            int dim_num0 = *(int *)DTSetGetNumberRange(&g_Element[i++], 2, 0, 8, "i");
            int reductionEnum[] = {0, 1, 2};
            int reduction = *(int *)DT_SetGetNumberEnum(&g_Element[i++], 0, reductionEnum, 3);
            int hasWeight = *(int *)DTSetGetNumberRange(&g_Element[i++], 0, 0, 1, "i");
            int hasPosWeight = *(int *)DTSetGetNumberRange(&g_Element[i++], 0, 0, 1, "i");

            std::vector<int64_t> selfShape = {};
            if (dim_num0 == 0) { selfShape = {}; } else 
            if (dim_num0 == 1) { selfShape = {shape1}; }
            else if (dim_num0 == 2) { selfShape = {shape1, shape2}; }
            else if (dim_num0 == 3) { selfShape = {shape1, shape2, shape3}; }
            else if (dim_num0 == 4) { selfShape = {shape1, shape2, shape3, shape4}; }
            else if (dim_num0 == 5) { selfShape = {shape1, shape2, shape3, shape4, shape5}; }
            else if (dim_num0 == 6) { selfShape = {shape1, shape2, shape3, shape4, shape5, shape6}; }
            else if (dim_num0 == 7) { selfShape = {shape1, shape2, shape3, shape4, shape5, shape6, shape7}; }
            else if (dim_num0 == 8) { selfShape = {shape1, shape2, shape3, shape4, shape5, shape6, shape7, shape8}; }

            std::vector<int64_t> targetShape = selfShape;
            std::vector<int64_t> outShape = (reduction == 0) ? selfShape : std::vector<int64_t>{1};

            long long shape_size = GetShapeSize(selfShape);
            if (shape_size > SHAPE_SIZE_CTL) { printf("shape_size too big\n"); goto ERROR; }
            long long outSize = GetShapeSize(outShape);

            if (dtype0 == aclDataType::ACL_FLOAT) {
                std::vector<float> selfHostData(shape_size, 1.0f);

                // === Tensor Size distribution check ===

                long long totalSize = shape_size + outSize;

                if (totalSize < totalSizeMin || totalSize > totalSizeMax) { goto ERROR; }

                ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, dtype0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                std::vector<float> targetHostData(shape_size, 1.0f);
                ret = CreateAclTensor(targetHostData, targetShape, &targetDeviceAddr, dtype0, &target);
                cstTargetShapeSz = (uint64_t)GetShapeSize(targetShape);
                cstTargetDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                if (hasWeight) {
                    std::vector<float> weightHostData(shape_size, 1.0f);
                    ret = CreateAclTensor(weightHostData, selfShape, &weightDeviceAddr, dtype0, &weight);
                    cstWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                if (hasPosWeight) {
                    std::vector<float> posWeightHostData(shape_size, 1.0f);
                    ret = CreateAclTensor(posWeightHostData, selfShape, &posWeightDeviceAddr, dtype0, &posWeight);
                    cstPosWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstPosWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                std::vector<float> outHostData(outSize, 0.0f);
                ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, dtype0, &out);
                cstOutShapeSz = (uint64_t)GetShapeSize(outShape);
                cstOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
            } else if (dtype0 == aclDataType::ACL_FLOAT16) {
                std::vector<int16_t> selfHostData(shape_size, 1);
                ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, dtype0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                std::vector<int16_t> targetHostData(shape_size, 1);
                ret = CreateAclTensor(targetHostData, targetShape, &targetDeviceAddr, dtype0, &target);
                cstTargetShapeSz = (uint64_t)GetShapeSize(targetShape);
                cstTargetDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                if (hasWeight) {
                    std::vector<int16_t> weightHostData(shape_size, 1);
                    ret = CreateAclTensor(weightHostData, selfShape, &weightDeviceAddr, dtype0, &weight);
                    cstWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                if (hasPosWeight) {
                    std::vector<int16_t> posWeightHostData(shape_size, 1);
                    ret = CreateAclTensor(posWeightHostData, selfShape, &posWeightDeviceAddr, dtype0, &posWeight);
                    cstPosWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstPosWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                std::vector<int16_t> outHostData(outSize, 0);
                ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, dtype0, &out);
                cstOutShapeSz = (uint64_t)GetShapeSize(outShape);
                cstOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
            } else if (dtype0 == aclDataType::ACL_BF16) {
                std::vector<int16_t> selfHostData(shape_size, 1);
                ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, dtype0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                std::vector<int16_t> targetHostData(shape_size, 1);
                ret = CreateAclTensor(targetHostData, targetShape, &targetDeviceAddr, dtype0, &target);
                cstTargetShapeSz = (uint64_t)GetShapeSize(targetShape);
                cstTargetDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                if (hasWeight) {
                    std::vector<int16_t> weightHostData(shape_size, 1);
                    ret = CreateAclTensor(weightHostData, selfShape, &weightDeviceAddr, dtype0, &weight);
                    cstWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                if (hasPosWeight) {
                    std::vector<int16_t> posWeightHostData(shape_size, 1);
                    ret = CreateAclTensor(posWeightHostData, selfShape, &posWeightDeviceAddr, dtype0, &posWeight);
                    cstPosWeightShapeSz = (uint64_t)GetShapeSize(selfShape);
                    cstPosWeightDtype = (int)(dtype0);
                    CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                }
                std::vector<int16_t> outHostData(outSize, 0);
                ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, dtype0, &out);
                cstOutShapeSz = (uint64_t)GetShapeSize(outShape);
                cstOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
            }

                        // ===== [CONSISTENT] 保存全部 tensor 设备侧初始内容（固定输入，含输入tensor的值） =====
            cstSelfBytes = (selfDeviceAddr != nullptr) ? (size_t)cstSelfShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstSelfDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=self, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstSelfDtype, ConsistentDtypeStr((aclDataType)cstSelfDtype), (unsigned long long)cstSelfShapeSz, cstSelfBytes, ConsistentSizeOfDtype((aclDataType)cstSelfDtype));
            if (cstSelfBytes == 0 && selfDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=self addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstSelfSaveBuf = (cstSelfBytes > 0) ? malloc(cstSelfBytes) : nullptr;
            if (cstSelfSaveBuf != nullptr && selfDeviceAddr != nullptr) {
                aclrtMemcpy(cstSelfSaveBuf, cstSelfBytes, selfDeviceAddr, cstSelfBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            cstTargetBytes = (targetDeviceAddr != nullptr) ? (size_t)cstTargetShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstTargetDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=target, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstTargetDtype, ConsistentDtypeStr((aclDataType)cstTargetDtype), (unsigned long long)cstTargetShapeSz, cstTargetBytes, ConsistentSizeOfDtype((aclDataType)cstTargetDtype));
            if (cstTargetBytes == 0 && targetDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=target addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstTargetSaveBuf = (cstTargetBytes > 0) ? malloc(cstTargetBytes) : nullptr;
            if (cstTargetSaveBuf != nullptr && targetDeviceAddr != nullptr) {
                aclrtMemcpy(cstTargetSaveBuf, cstTargetBytes, targetDeviceAddr, cstTargetBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            cstWeightBytes = (weightDeviceAddr != nullptr) ? (size_t)cstWeightShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstWeightDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=weight, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstWeightDtype, ConsistentDtypeStr((aclDataType)cstWeightDtype), (unsigned long long)cstWeightShapeSz, cstWeightBytes, ConsistentSizeOfDtype((aclDataType)cstWeightDtype));
            if (cstWeightBytes == 0 && weightDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=weight addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstWeightSaveBuf = (cstWeightBytes > 0) ? malloc(cstWeightBytes) : nullptr;
            if (cstWeightSaveBuf != nullptr && weightDeviceAddr != nullptr) {
                aclrtMemcpy(cstWeightSaveBuf, cstWeightBytes, weightDeviceAddr, cstWeightBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            cstPosWeightBytes = (posWeightDeviceAddr != nullptr) ? (size_t)cstPosWeightShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstPosWeightDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=posWeight, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstPosWeightDtype, ConsistentDtypeStr((aclDataType)cstPosWeightDtype), (unsigned long long)cstPosWeightShapeSz, cstPosWeightBytes, ConsistentSizeOfDtype((aclDataType)cstPosWeightDtype));
            if (cstPosWeightBytes == 0 && posWeightDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=posWeight addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstPosWeightSaveBuf = (cstPosWeightBytes > 0) ? malloc(cstPosWeightBytes) : nullptr;
            if (cstPosWeightSaveBuf != nullptr && posWeightDeviceAddr != nullptr) {
                aclrtMemcpy(cstPosWeightSaveBuf, cstPosWeightBytes, posWeightDeviceAddr, cstPosWeightBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            cstOutBytes = (outDeviceAddr != nullptr) ? (size_t)cstOutShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstOutDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=out, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstOutDtype, ConsistentDtypeStr((aclDataType)cstOutDtype), (unsigned long long)cstOutShapeSz, cstOutBytes, ConsistentSizeOfDtype((aclDataType)cstOutDtype));
            if (cstOutBytes == 0 && outDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=out addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstOutSaveBuf = (cstOutBytes > 0) ? malloc(cstOutBytes) : nullptr;
            if (cstOutSaveBuf != nullptr && outDeviceAddr != nullptr) {
                aclrtMemcpy(cstOutSaveBuf, cstOutBytes, outDeviceAddr, cstOutBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }

            // ===== [CONSISTENT] 固定输入重复执行 CONSISTENT_RUN_TIMES 次，逐轮对全部 tensor 计算 hash =====
            cstRunCount = 0;
            cstAllOk = true;
            for (int cstRun = 0; cstRun < CONSISTENT_RUN_TIMES; cstRun++) {
                // [CONSISTENT] 恢复固定输入（输出 tensor 同时恢复初始值，覆盖 inplace/累加类算子）
                if (cstSelfSaveBuf != nullptr && selfDeviceAddr != nullptr) {
                    aclrtMemcpy(selfDeviceAddr, cstSelfBytes, cstSelfSaveBuf, cstSelfBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstTargetSaveBuf != nullptr && targetDeviceAddr != nullptr) {
                    aclrtMemcpy(targetDeviceAddr, cstTargetBytes, cstTargetSaveBuf, cstTargetBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstWeightSaveBuf != nullptr && weightDeviceAddr != nullptr) {
                    aclrtMemcpy(weightDeviceAddr, cstWeightBytes, cstWeightSaveBuf, cstWeightBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstPosWeightSaveBuf != nullptr && posWeightDeviceAddr != nullptr) {
                    aclrtMemcpy(posWeightDeviceAddr, cstPosWeightBytes, cstPosWeightSaveBuf, cstPosWeightBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstOutSaveBuf != nullptr && outDeviceAddr != nullptr) {
                    aclrtMemcpy(outDeviceAddr, cstOutBytes, cstOutSaveBuf, cstOutBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                ret = aclnnBinaryCrossEntropyWithLogitsGetWorkspaceSize(self, target, weight, posWeight, reduction, out, &workspaceSize, &executor);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnBinaryCrossEntropyWithLogits execute failed :%d\n", (int)ret);
                    printf("[FUZZ_CONSISTENT] case_%d: faild, type=execute_error, stage=GetWorkspaceSize, run=%d/%d, ret=%d\n", fuzzi, cstRun + 1, CONSISTENT_RUN_TIMES, (int)ret);
                    cstAllOk = false;
                    break;
                }
                if (workspaceSize > 0) {
                    if (workspaceAddr != nullptr) { aclrtFree(workspaceAddr); workspaceAddr = nullptr; }
                    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
                    if (ret != ACL_SUCCESS) {
                        printf("[FUZZ_CONSISTENT] case_%d: faild, type=execute_error, stage=workspace_malloc, run=%d/%d, ret=%d\n", fuzzi, cstRun + 1, CONSISTENT_RUN_TIMES, (int)ret);
                        cstAllOk = false;
                        break;
                    }
                }
                ret = aclnnBinaryCrossEntropyWithLogits(workspaceAddr, workspaceSize, executor, stream);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnBinaryCrossEntropyWithLogits execute failed :%d\n", (int)ret);
                    printf("[FUZZ_CONSISTENT] case_%d: faild, type=execute_error, stage=execute, run=%d/%d, ret=%d\n", fuzzi, cstRun + 1, CONSISTENT_RUN_TIMES, (int)ret);
                    cstAllOk = false;
                    break;
                }
                ret = aclrtSynchronizeStream(stream);
                if (ret != ACL_SUCCESS) {
                    printf("[FUZZ_CONSISTENT] case_%d: faild, type=execute_error, stage=sync, run=%d/%d, ret=%d\n", fuzzi, cstRun + 1, CONSISTENT_RUN_TIMES, (int)ret);
                    cstAllOk = false;
                    break;
                }
                // [CONSISTENT] 本轮执行结果 hash
                if (selfDeviceAddr != nullptr && cstSelfBytes > 0) {
                    cstSelfHashArr[cstRun] = ConsistentHashDeviceTensor(selfDeviceAddr, cstSelfBytes);
                }
                if (targetDeviceAddr != nullptr && cstTargetBytes > 0) {
                    cstTargetHashArr[cstRun] = ConsistentHashDeviceTensor(targetDeviceAddr, cstTargetBytes);
                }
                if (weightDeviceAddr != nullptr && cstWeightBytes > 0) {
                    cstWeightHashArr[cstRun] = ConsistentHashDeviceTensor(weightDeviceAddr, cstWeightBytes);
                }
                if (posWeightDeviceAddr != nullptr && cstPosWeightBytes > 0) {
                    cstPosWeightHashArr[cstRun] = ConsistentHashDeviceTensor(posWeightDeviceAddr, cstPosWeightBytes);
                }
                if (outDeviceAddr != nullptr && cstOutBytes > 0) {
                    cstOutHashArr[cstRun] = ConsistentHashDeviceTensor(outDeviceAddr, cstOutBytes);
                }
                cstRunCount++;
                printf("aclnnBinaryCrossEntropyWithLogits execute success :%d\n", (int)ret);
            }

            // ===== [CONSISTENT] 对比 2 轮 hash，判定一致性（2次执行输出一致即为确定性） =====
            if (cstAllOk && cstRunCount == CONSISTENT_RUN_TIMES) {
                bool cstSelfSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstSelfHashArr[cstR] != cstSelfHashArr[0]) { cstSelfSame = false; break; }
                }
                bool cstTargetSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstTargetHashArr[cstR] != cstTargetHashArr[0]) { cstTargetSame = false; break; }
                }
                bool cstWeightSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstWeightHashArr[cstR] != cstWeightHashArr[0]) { cstWeightSame = false; break; }
                }
                bool cstPosWeightSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstPosWeightHashArr[cstR] != cstPosWeightHashArr[0]) { cstPosWeightSame = false; break; }
                }
                bool cstOutSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstOutHashArr[cstR] != cstOutHashArr[0]) { cstOutSame = false; break; }
                }
                bool cstAllSame = cstSelfSame && cstTargetSame && cstWeightSame && cstPosWeightSame && cstOutSame;
                if (cstAllSame) {
                    printf("[FUZZ_CONSISTENT] case_%d: success, type=consistent, runs=%d, hash_consistent=all\n", fuzzi, CONSISTENT_RUN_TIMES);
                    printf("[FUZZ_RESULT] case_%d: success, ret=%d\n", fuzzi, (int)ret);
                    cstResultPrinted = true;
                } else {
                    printf("[FUZZ_CONSISTENT] case_%d: faild, type=hash_mismatch, runs=%d, inconsistent=", fuzzi, CONSISTENT_RUN_TIMES);
                    if (!cstSelfSame) printf("self ");
                    if (!cstTargetSame) printf("target ");
                    if (!cstWeightSame) printf("weight ");
                    if (!cstPosWeightSame) printf("posWeight ");
                    if (!cstOutSame) printf("out ");
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: self hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstSelfHashArr[cstR]);
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: target hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstTargetHashArr[cstR]);
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: weight hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstWeightHashArr[cstR]);
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: posWeight hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstPosWeightHashArr[cstR]);
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: out hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstOutHashArr[cstR]);
                    printf("\n");
                    // [CONSISTENT] 失败时打印全部入参信息（dtype/format/shape/attr/tiling_key）
                    printf("[FUZZ_PARAM] ");
                    printf("self: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstSelfDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                    printf("target: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstTargetDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstTargetShapeSz); printf("; ");
                    printf("weight: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstWeightDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstWeightShapeSz); printf("; ");
                    printf("posWeight: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstPosWeightDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstPosWeightShapeSz); printf("; ");
                    printf("out: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstOutDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstOutShapeSz); printf("; ");
                    printf("reduction=%d; ", (int)reduction);
                    printf("tiling_key=%s\n", curTilingKey);
                    // ret=-2 为 hash_mismatch 专用标记（非 aclnn 错误码：执行成功但输出不确定）
                    printf("[FUZZ_RESULT] case_%d: faild, ret=%d\n", fuzzi, -2);
                    cstResultPrinted = true;
                }
            } else if (!cstAllOk) {
                // [CONSISTENT] 执行失败也打印全部入参信息
                printf("[FUZZ_PARAM] ");
                printf("self: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstSelfDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                printf("target: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstTargetDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstTargetShapeSz); printf("; ");
                printf("weight: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstWeightDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstWeightShapeSz); printf("; ");
                printf("posWeight: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstPosWeightDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstPosWeightShapeSz); printf("; ");
                printf("out: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstOutDtype), ConsistentFormatStr(ACL_FORMAT_ND)); printf("shapeSize=%llu", (unsigned long long)cstOutShapeSz); printf("; ");
                printf("reduction=%d; ", (int)reduction);
                printf("tiling_key=%s\n", curTilingKey);
                printf("[FUZZ_RESULT] case_%d: faild, ret=%d\n", fuzzi, (int)ret);
                cstResultPrinted = true;
            }
        }
        catch (std::exception& e)
        {
            printf("catch exception %s\n", e.what());
            printf("[FUZZ_CONSISTENT] case_%d: faild, type=exception, ret=%d\n", fuzzi, (int)ret);
            printf("[FUZZ_RESULT] case_%d: faild, ret=%d\n", fuzzi, (int)ret);
            cstResultPrinted = true;
        }

ERROR:
        printf("end \n");
        CheckAndLog(ret);
        pid_t pid = getpid();
        std::string cmd = "grep -rin \'aicore error\' /root/ascend/log/debug/plog/ ";
        printf("system return value %d\n", CheckResult(pid));
        // [CONSISTENT] 释放固定输入快照
        if (cstSelfSaveBuf != nullptr) { free(cstSelfSaveBuf); cstSelfSaveBuf = nullptr; }
        if (cstTargetSaveBuf != nullptr) { free(cstTargetSaveBuf); cstTargetSaveBuf = nullptr; }
        if (cstWeightSaveBuf != nullptr) { free(cstWeightSaveBuf); cstWeightSaveBuf = nullptr; }
        if (cstPosWeightSaveBuf != nullptr) { free(cstPosWeightSaveBuf); cstPosWeightSaveBuf = nullptr; }
        if (cstOutSaveBuf != nullptr) { free(cstOutSaveBuf); cstOutSaveBuf = nullptr; }
        aclDestroyTensor(self);
        aclDestroyTensor(target);
        aclDestroyTensor(weight);
        aclDestroyTensor(posWeight);
        aclDestroyTensor(out);
        aclrtFree(selfDeviceAddr);
        aclrtFree(targetDeviceAddr);
        aclrtFree(weightDeviceAddr);
        aclrtFree(posWeightDeviceAddr);
        aclrtFree(outDeviceAddr);
        if (workspaceSize > 0) { aclrtFree(workspaceAddr); }
    }
    DT_FUZZ_END()
    WriteDTLog("aclnnBinaryCrossEntropyWithLogits", "aclnnBinaryCrossEntropyWithLogits consistent op fuzz total number:%u, para error number %u, execute error number %u \n", fuzz_num, aclnn_INVALID_Para, aclnn_Inner_Error + Runtime_Error);
    WriteDTLog("aclnnBinaryCrossEntropyWithLogits", "aclnnBinaryCrossEntropyWithLogits consistent op fuzz end \n");
    aclrtDestroyStream(stream);
    aclrtDestroyContext(context);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}

extern "C" int aclnnBinaryCrossEntropyWithLogits_consistent_fuzz(int argc, char *argv[]);
int aclnnBinaryCrossEntropyWithLogits_consistent_fuzz(int argc, char *argv[])
{
    // [CONSISTENT] 运行时覆盖每用例执行次数（环境变量 CONSISTENT_RUN_TIMES：默认2次，可按需覆盖）
    const char* cstRunEnv = getenv("CONSISTENT_RUN_TIMES");
    if (cstRunEnv != nullptr) {
        int cstRunV = atoi(cstRunEnv);
        if (cstRunV > 0 && cstRunV <= CONSISTENT_MAX_RUN_TIMES) { CONSISTENT_RUN_TIMES = cstRunV; }
    }
    printf("[FUZZ_CONSISTENT_RUNTIMES] run_times=%d\n", CONSISTENT_RUN_TIMES);
    int fuzz_num = MAXFUZZNUMBER;
    if (argc < 1) {
        cout << "[ERROR] input arg num must be bigger than 1" << endl;
        return -1;
    } else if (argc > 1) {
        fuzz_num = atoi(argv[2]);
    }
    if (argc == 4) { deviceId = atoi(argv[3]); }
    printf("fuzz_num %d \n", fuzz_num);
    printf("aclnnBinaryCrossEntropyWithLogits_consistent :%d \n", aclnnBinaryCrossEntropyWithLogits_consistent(fuzz_num));
    return 0;
}