#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_avgpool2d_backward.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 ========== */

static int kDtypeEnum[] = {0, 1, 27};
static int kDtypeNum = sizeof(kDtypeEnum) / sizeof(kDtypeEnum[0]);

int aclnnAvgPool2dBackward_consistent(int fuzz_num)
{
    const long long kSingleInputSizeMax = 1073741824LL;

    auto ret = Init(deviceId, &context, &stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // [CONSISTENT] md文档标注非确定性实现，打开确定性计算开关
    ret = aclrtCtxSetSysParamOpt(aclSysParamOpt::ACL_OPT_DETERMINISTIC, 1);
    printf("[FUZZ_DETERMINISTIC] api=aclnnAvgPool2dBackward, enable_deterministic=1, ret=%d\n", (int)ret);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCtxSetSysParamOpt failed. ERROR: %d\n", ret); return ret);
    WriteDTLog("aclnnAvgPool2dBackward", "aclnnAvgPool2dBackward consistent op fuzz start \n");

    aclrtSetOpExecuteTimeOut(1091);
    DT_Set_TimeOut_Second(3000);
    DT_FUZZ_START(0, fuzz_num, (char*)__func__, 0)
    {
        int i = 0;

        void* gradOutputDeviceAddr = nullptr;
        aclTensor* gradOutput = nullptr;
        void* selfDeviceAddr = nullptr;
        aclTensor* self = nullptr;
        void* gradInputDeviceAddr = nullptr;
        aclTensor* gradInput = nullptr;
        aclIntArray* kernelSize = nullptr;
        aclIntArray* stride = nullptr;
        aclIntArray* padding = nullptr;

        uint64_t workspaceSize = 0;
        aclOpExecutor* executor;
        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 cstGradOutputShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstGradOutputDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstGradOutputBytes = 0;
        void* cstGradOutputSaveBuf = nullptr;
        uint64_t cstGradOutputHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstGradOutputHashArr, 0, sizeof(cstGradOutputHashArr));
        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 cstGradInputShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstGradInputDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstGradInputBytes = 0;
        void* cstGradInputSaveBuf = nullptr;
        uint64_t cstGradInputHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstGradInputHashArr, 0, sizeof(cstGradInputHashArr));

        printf("[FUZZ_LOG] plog_file: aclnnAvgPool2dBackward_consistent_plog_case_%d\n", fuzzi);
        LOG_PRINT("aclnnAvgPool2dBackward_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 N = *(int*)DTSetGetNumberRange(&g_Element[i++], 2, 1, 4, "i");
            int C = *(int*)DTSetGetNumberRange(&g_Element[i++], 2, 1, 4, "i");
            int H = *(int*)DTSetGetNumberRange(&g_Element[i++], 10, 4, 16, "i");
            int W = *(int*)DTSetGetNumberRange(&g_Element[i++], 10, 4, 16, "i");

            auto selfDtype = (aclDataType)*(int*)DT_SetGetNumberEnumV3(i++, aclDataType::ACL_FLOAT, kDtypeEnum, kDtypeNum);

            if (N < 1) N = 1; if (N > 4) N = 4;
            if (C < 1) C = 1; if (C > 4) C = 4;
            if (H < 4) H = 4; if (H > 16) H = 16;
            if (W < 4) W = 4; if (W > 16) W = 16;

            int64_t oH = (int64_t)H - 1;
            int64_t oW = (int64_t)W - 1;
            if (oH < 1) oH = 1;
            if (oW < 1) oW = 1;

            std::vector<int64_t> gradOutputShape = {(int64_t)N, (int64_t)C, oH, oW};
            std::vector<int64_t> selfShape = {(int64_t)N, (int64_t)C, (int64_t)H, (int64_t)W};
            std::vector<int64_t> gradInputShape = selfShape;

            int64_t ksArr[] = {2, 2};
            kernelSize = aclCreateIntArray(ksArr, 2);
            CHECK_RET(kernelSize != nullptr, LOG_PRINT("aclCreateIntArray kernelSize failed.\n"); goto ERROR);
            int64_t strideArr[] = {1, 1};
            stride = aclCreateIntArray(strideArr, 2);
            CHECK_RET(stride != nullptr, LOG_PRINT("aclCreateIntArray stride failed.\n"); goto ERROR);
            int64_t padArr[] = {0, 0};
            padding = aclCreateIntArray(padArr, 2);
            CHECK_RET(padding != nullptr, LOG_PRINT("aclCreateIntArray padding failed.\n"); goto ERROR);

            bool ceilMode = false;
            bool countIncludePad = true;
            int64_t divisorOverride = 0;
            int8_t cubeMathType = 0;

            long long gradOutputSize = GetShapeSize(gradOutputShape);
            long long selfSize = GetShapeSize(selfShape);
            long long gradInputSize = GetShapeSize(gradInputShape);
            if (gradOutputSize > kSingleInputSizeMax || selfSize > kSingleInputSizeMax || gradInputSize > kSingleInputSizeMax) { printf("shape_size too big\n"); goto ERROR; }

            printf("selfDtype=%d, N=%d, C=%d, H=%d, W=%d, oH=%lld, oW=%lld\n", (int)selfDtype, N, C, H, W, (long long)oH, (long long)oW);

            aclFormat fmt = aclFormat::ACL_FORMAT_NCHW;

            if (selfDtype == aclDataType::ACL_FLOAT) {
                std::vector<float> hostData(gradOutputSize, 1);

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

                long long totalSize = gradOutputSize + selfSize + gradInputSize;

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

                ret = CreateAclTensor_New(hostData, gradOutputShape, &gradOutputDeviceAddr, selfDtype, fmt, &gradOutput);
                cstGradOutputShapeSz = (uint64_t)GetShapeSize(gradOutputShape);
                cstGradOutputDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_FLOAT16) {
                std::vector<int16_t> hostData(gradOutputSize, 1);
                ret = CreateAclTensor_New(hostData, gradOutputShape, &gradOutputDeviceAddr, selfDtype, fmt, &gradOutput);
                cstGradOutputShapeSz = (uint64_t)GetShapeSize(gradOutputShape);
                cstGradOutputDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_BF16) {
                std::vector<int16_t> hostData(gradOutputSize, 1);
                ret = CreateAclTensor_New(hostData, gradOutputShape, &gradOutputDeviceAddr, selfDtype, fmt, &gradOutput);
                cstGradOutputShapeSz = (uint64_t)GetShapeSize(gradOutputShape);
                cstGradOutputDtype = (int)(selfDtype);
            }
            CHECK_RET(ret == ACL_SUCCESS, goto ERROR);

            if (selfDtype == aclDataType::ACL_FLOAT) {
                std::vector<float> hostData(selfSize, 1);
                ret = CreateAclTensor_New(hostData, selfShape, &selfDeviceAddr, selfDtype, fmt, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_FLOAT16) {
                std::vector<int16_t> hostData(selfSize, 1);
                ret = CreateAclTensor_New(hostData, selfShape, &selfDeviceAddr, selfDtype, fmt, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_BF16) {
                std::vector<int16_t> hostData(selfSize, 1);
                ret = CreateAclTensor_New(hostData, selfShape, &selfDeviceAddr, selfDtype, fmt, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(selfDtype);
            }
            CHECK_RET(ret == ACL_SUCCESS, goto ERROR);

            if (selfDtype == aclDataType::ACL_FLOAT) {
                std::vector<float> hostData(gradInputSize, 0);
                ret = CreateAclTensor_New(hostData, gradInputShape, &gradInputDeviceAddr, selfDtype, fmt, &gradInput);
                cstGradInputShapeSz = (uint64_t)GetShapeSize(gradInputShape);
                cstGradInputDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_FLOAT16) {
                std::vector<int16_t> hostData(gradInputSize, 0);
                ret = CreateAclTensor_New(hostData, gradInputShape, &gradInputDeviceAddr, selfDtype, fmt, &gradInput);
                cstGradInputShapeSz = (uint64_t)GetShapeSize(gradInputShape);
                cstGradInputDtype = (int)(selfDtype);
            } else if (selfDtype == aclDataType::ACL_BF16) {
                std::vector<int16_t> hostData(gradInputSize, 0);
                ret = CreateAclTensor_New(hostData, gradInputShape, &gradInputDeviceAddr, selfDtype, fmt, &gradInput);
                cstGradInputShapeSz = (uint64_t)GetShapeSize(gradInputShape);
                cstGradInputDtype = (int)(selfDtype);
            }
            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） =====
            cstGradOutputBytes = (gradOutputDeviceAddr != nullptr) ? (size_t)cstGradOutputShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstGradOutputDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=gradOutput, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstGradOutputDtype, ConsistentDtypeStr((aclDataType)cstGradOutputDtype), (unsigned long long)cstGradOutputShapeSz, cstGradOutputBytes, ConsistentSizeOfDtype((aclDataType)cstGradOutputDtype));
            if (cstGradOutputBytes == 0 && gradOutputDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=gradOutput addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstGradOutputSaveBuf = (cstGradOutputBytes > 0) ? malloc(cstGradOutputBytes) : nullptr;
            if (cstGradOutputSaveBuf != nullptr && gradOutputDeviceAddr != nullptr) {
                aclrtMemcpy(cstGradOutputSaveBuf, cstGradOutputBytes, gradOutputDeviceAddr, cstGradOutputBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            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);
            }
            cstGradInputBytes = (gradInputDeviceAddr != nullptr) ? (size_t)cstGradInputShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstGradInputDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=gradInput, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstGradInputDtype, ConsistentDtypeStr((aclDataType)cstGradInputDtype), (unsigned long long)cstGradInputShapeSz, cstGradInputBytes, ConsistentSizeOfDtype((aclDataType)cstGradInputDtype));
            if (cstGradInputBytes == 0 && gradInputDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=gradInput addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstGradInputSaveBuf = (cstGradInputBytes > 0) ? malloc(cstGradInputBytes) : nullptr;
            if (cstGradInputSaveBuf != nullptr && gradInputDeviceAddr != nullptr) {
                aclrtMemcpy(cstGradInputSaveBuf, cstGradInputBytes, gradInputDeviceAddr, cstGradInputBytes, 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 (cstGradOutputSaveBuf != nullptr && gradOutputDeviceAddr != nullptr) {
                    aclrtMemcpy(gradOutputDeviceAddr, cstGradOutputBytes, cstGradOutputSaveBuf, cstGradOutputBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstSelfSaveBuf != nullptr && selfDeviceAddr != nullptr) {
                    aclrtMemcpy(selfDeviceAddr, cstSelfBytes, cstSelfSaveBuf, cstSelfBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstGradInputSaveBuf != nullptr && gradInputDeviceAddr != nullptr) {
                    aclrtMemcpy(gradInputDeviceAddr, cstGradInputBytes, cstGradInputSaveBuf, cstGradInputBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                ret = aclnnAvgPool2dBackwardGetWorkspaceSize(gradOutput, self, kernelSize, stride, padding, ceilMode, countIncludePad, divisorOverride, cubeMathType, gradInput, &workspaceSize, &executor);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnAvgPool2dBackward 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 = aclnnAvgPool2dBackward(workspaceAddr, workspaceSize, executor, stream);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnAvgPool2dBackward 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 (gradOutputDeviceAddr != nullptr && cstGradOutputBytes > 0) {
                    cstGradOutputHashArr[cstRun] = ConsistentHashDeviceTensor(gradOutputDeviceAddr, cstGradOutputBytes);
                }
                if (selfDeviceAddr != nullptr && cstSelfBytes > 0) {
                    cstSelfHashArr[cstRun] = ConsistentHashDeviceTensor(selfDeviceAddr, cstSelfBytes);
                }
                if (gradInputDeviceAddr != nullptr && cstGradInputBytes > 0) {
                    cstGradInputHashArr[cstRun] = ConsistentHashDeviceTensor(gradInputDeviceAddr, cstGradInputBytes);
                }
                cstRunCount++;
                printf("aclnnAvgPool2dBackward 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 cstGradOutputSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstGradOutputHashArr[cstR] != cstGradOutputHashArr[0]) { cstGradOutputSame = false; break; }
                }
                bool cstSelfSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstSelfHashArr[cstR] != cstSelfHashArr[0]) { cstSelfSame = false; break; }
                }
                bool cstGradInputSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstGradInputHashArr[cstR] != cstGradInputHashArr[0]) { cstGradInputSame = false; break; }
                }
                bool cstAllSame = cstGradOutputSame && cstSelfSame && cstGradInputSame;
                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 (!cstGradOutputSame) printf("gradOutput ");
                    if (!cstSelfSame) printf("self ");
                    if (!cstGradInputSame) printf("gradInput ");
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: gradOutput hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstGradOutputHashArr[cstR]);
                    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: gradInput hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstGradInputHashArr[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("gradOutput: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstGradOutputDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstGradOutputShapeSz); printf("; ");
                    printf("self: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstSelfDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                    printf("gradInput: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstGradInputDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstGradInputShapeSz); printf("; ");
                    printf("kernelSize: len=%zu; ", (size_t)((size_t)2));
                    printf("stride: len=%zu; ", (size_t)((size_t)2));
                    printf("padding: len=%zu; ", (size_t)((size_t)2));
                    printf("ceilMode=%d; ", (int)((ceilMode) ? 1 : 0));
                    printf("countIncludePad=%d; ", (int)((countIncludePad) ? 1 : 0));
                    printf("divisorOverride=%lld; ", (long long)divisorOverride);
                    printf("cubeMathType=%d; ", (int)cubeMathType);
                    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("gradOutput: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstGradOutputDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstGradOutputShapeSz); printf("; ");
                printf("self: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstSelfDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                printf("gradInput: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstGradInputDtype), ConsistentFormatStr(fmt)); printf("shapeSize=%llu", (unsigned long long)cstGradInputShapeSz); printf("; ");
                printf("kernelSize: len=%zu; ", (size_t)((size_t)2));
                printf("stride: len=%zu; ", (size_t)((size_t)2));
                printf("padding: len=%zu; ", (size_t)((size_t)2));
                printf("ceilMode=%d; ", (int)((ceilMode) ? 1 : 0));
                printf("countIncludePad=%d; ", (int)((countIncludePad) ? 1 : 0));
                printf("divisorOverride=%lld; ", (long long)divisorOverride);
                printf("cubeMathType=%d; ", (int)cubeMathType);
                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));

        if (kernelSize) { aclDestroyIntArray(kernelSize); }
        if (stride) { aclDestroyIntArray(stride); }
        if (padding) { aclDestroyIntArray(padding); }
        // [CONSISTENT] 释放固定输入快照
        if (cstGradOutputSaveBuf != nullptr) { free(cstGradOutputSaveBuf); cstGradOutputSaveBuf = nullptr; }
        if (cstSelfSaveBuf != nullptr) { free(cstSelfSaveBuf); cstSelfSaveBuf = nullptr; }
        if (cstGradInputSaveBuf != nullptr) { free(cstGradInputSaveBuf); cstGradInputSaveBuf = nullptr; }
        aclDestroyTensor(gradOutput);
        aclDestroyTensor(self);
        aclDestroyTensor(gradInput);
        aclrtFree(gradOutputDeviceAddr);
        aclrtFree(selfDeviceAddr);
        aclrtFree(gradInputDeviceAddr);
        if (workspaceSize > 0 && workspaceAddr) { aclrtFree(workspaceAddr); }
    }
    DT_FUZZ_END()

    WriteDTLog("aclnnAvgPool2dBackward", "aclnnAvgPool2dBackward 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("aclnnAvgPool2dBackward", "aclnnAvgPool2dBackward consistent op fuzz end \n");
    aclrtDestroyStream(stream);
    aclrtDestroyContext(context);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}

extern "C" int aclnnAvgPool2dBackward_consistent_fuzz(int argc, char *argv[]);
int aclnnAvgPool2dBackward_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 \n" << 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("aclnnAvgPool2dBackward_consistent :%d \n", aclnnAvgPool2dBackward_consistent(fuzz_num));
    return 0;
}