#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 <unistd.h>
#include <cstdio>
#include <string>
#include <vector>
#include <iostream>
#include <iomanip>
#include <string.h>
#include <time.h>
#include <ctime>
#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 <complex>
#include "../fuzzlib/secodeFuzz.h"
#include "aclnnop/aclnn_aminmax_all.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 aclnnAminmaxAll_consistent(int fuzz_num)
{
    auto ret = Init(deviceId, &context, &stream);
    // check根据自己的需要处理
    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=aclnnAminmaxAll, enable_deterministic=0, reason=md文档标注：aclnnAminmaxAll默认确定性实现。\n");
    WriteDTLog("aclnnAminmaxAll", "aclnnAminmaxAll consistent op fuzz start \n");
	// aclDataType definition at acl/acl_base.h
    int dtype0Enum[] = { 0, 27, 1, 11, 2, 6, 3, 9, 4, 12 };
    int dtype0Num = sizeof(dtype0Enum) / sizeof(dtype0Enum[0]);
	int dim0Enum[] = { 1, 2, 3, 4, 5, 6, 7, 8 };
    int dim0Num = sizeof(dim0Enum) / sizeof(dim0Enum[0]);

    aclrtSetOpExecuteTimeOut(1091);
    DT_Set_TimeOut_Second(3000);
    DT_FUZZ_START(0, fuzz_num, "aclnnAminmaxAll_consistent",0)
    {
        int i=0;
		void* selfDeviceAddr = nullptr;
		void* maxOutDeviceAddr = nullptr;
		void* minOutDeviceAddr = nullptr;
		aclTensor* self = nullptr;
		aclTensor* maxOut = nullptr;
		aclTensor* minOut = 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 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 cstMaxOutShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstMaxOutDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstMaxOutBytes = 0;
        void* cstMaxOutSaveBuf = nullptr;
        uint64_t cstMaxOutHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstMaxOutHashArr, 0, sizeof(cstMaxOutHashArr));
        uint64_t cstMinOutShapeSz = 0;  // 创建语句处捕获的 shape size（作用域安全）
        int cstMinOutDtype = 0;  // 创建语句处捕获的 dtype（作用域安全）
        size_t cstMinOutBytes = 0;
        void* cstMinOutSaveBuf = nullptr;
        uint64_t cstMinOutHashArr[CONSISTENT_MAX_RUN_TIMES];
        memset(cstMinOutHashArr, 0, sizeof(cstMinOutHashArr));

        printf("[FUZZ_LOG] plog_file: aclnnAminmaxAll_consistent_plog_case_%d\n", fuzzi);
        LOG_PRINT("aclnnAminmaxAll_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;

		    }

        // 创建调用DT获取变异数据
			auto dtype0 = (aclDataType)*(int *)DT_SetGetNumberEnumV3(i++, aclDataType::ACL_FLOAT, dtype0Enum, dtype0Num);
            printf("dtype0: %d\n",dtype0);
			int dim_num0 = *(int *)DT_SetGetNumberEnumV3(i++, 1, dim0Enum, dim0Num);

			std::vector <int64_t> selfShape = {};
			std::vector <int64_t> maxOutShape = {};
			std::vector <int64_t> minOutShape = {};
			for (int64_t index = 0; index < dim_num0; index++){
                int64_t dim_value = *(int * )DTSetGetNumberRange(&g_Element[i++], 5, 1, MAX_SHAPE_SIZE, "i");
				selfShape.push_back(dim_value);
			}
			for (int64_t index = 0; index < dim_num0; index++){
                // int64_t dim_value = *(int * )DTSetGetNumberRange(&g_Element[i++], 5, 1, MAX_SHAPE_SIZE, "i");
                int dim_value=1;
				maxOutShape.push_back(dim_value);
			}
			for (int64_t index = 0; index < dim_num0; index++){
                // int64_t dim_value = *(int * )DTSetGetNumberRange(&g_Element[i++], 5, 1, MAX_SHAPE_SIZE, "i");
                int dim_value=1;
				minOutShape.push_back(dim_value);
			}
			aclFormat format0;
			if(dim_num0 == 1){
				printf("selfShape:{%d}\n",selfShape[0]);
				printf("maxOutShape:{%d}\n",maxOutShape[0]);
				printf("minOutShape:{%d}\n",minOutShape[0]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 2){
				printf("selfShape:{%d,%d}\n",selfShape[0],selfShape[1]);
				printf("maxOutShape:{%d,%d}\n",maxOutShape[0],maxOutShape[1]);
				printf("minOutShape:{%d,%d}\n",minOutShape[0],minOutShape[1]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 3){
				printf("selfShape:{%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2]);
				printf("maxOutShape:{%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2]);
				printf("minOutShape:{%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 4){
				printf("selfShape:{%d,%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2],selfShape[3]);
				printf("maxOutShape:{%d,%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2],maxOutShape[3]);
				printf("minOutShape:{%d,%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2],minOutShape[3]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 5){
				printf("selfShape:{%d,%d,%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2],selfShape[3],selfShape[4]);
				printf("maxOutShape:{%d,%d,%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2],maxOutShape[3],maxOutShape[4]);
				printf("minOutShape:{%d,%d,%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2],minOutShape[3],minOutShape[4]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 6){
				printf("selfShape:{%d,%d,%d,%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2],selfShape[3],selfShape[4],selfShape[5]);
				printf("maxOutShape:{%d,%d,%d,%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2],maxOutShape[3],maxOutShape[4],maxOutShape[5]);
				printf("minOutShape:{%d,%d,%d,%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2],minOutShape[3],minOutShape[4],minOutShape[5]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 7){
				printf("selfShape:{%d,%d,%d,%d,%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2],selfShape[3],selfShape[4],selfShape[5],selfShape[6]);
				printf("maxOutShape:{%d,%d,%d,%d,%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2],maxOutShape[3],maxOutShape[4],maxOutShape[5],maxOutShape[6]);
				printf("minOutShape:{%d,%d,%d,%d,%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2],minOutShape[3],minOutShape[4],minOutShape[5],minOutShape[6]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} else if(dim_num0 == 8){
				printf("selfShape:{%d,%d,%d,%d,%d,%d,%d,%d}\n",selfShape[0],selfShape[1],selfShape[2],selfShape[3],selfShape[4],selfShape[5],selfShape[6],selfShape[7]);
				printf("maxOutShape:{%d,%d,%d,%d,%d,%d,%d,%d}\n",maxOutShape[0],maxOutShape[1],maxOutShape[2],maxOutShape[3],maxOutShape[4],maxOutShape[5],maxOutShape[6],maxOutShape[7]);
				printf("minOutShape:{%d,%d,%d,%d,%d,%d,%d,%d}\n",minOutShape[0],minOutShape[1],minOutShape[2],minOutShape[3],minOutShape[4],minOutShape[5],minOutShape[6],minOutShape[7]);
				format0 = aclFormat::ACL_FORMAT_ND;
			} 
			long long selfShape_size = GetShapeSize(selfShape);
            if(selfShape_size > SHAPE_SIZE_CTL)
            {
                printf("selfShape_size too big\n");
                goto ERROR;
            }
			long long maxOutShape_size = GetShapeSize(maxOutShape);
            if(maxOutShape_size > SHAPE_SIZE_CTL)
            {
                printf("maxOutShape_size too big\n");
                goto ERROR;
            }
			long long minOutShape_size = GetShapeSize(minOutShape);
            if(minOutShape_size > SHAPE_SIZE_CTL)
            {
                printf("minOutShape_size too big\n");
                goto ERROR;
            }
			if (dtype0 == aclDataType::ACL_FLOAT) {
				// 创建self aclTensor
                std::vector<float> selfHostData (selfShape_size , 1);

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

                long long totalSize = selfShape_size + maxOutShape_size + minOutShape_size;

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

                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<float> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<float> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_BF16) {
				// 创建self aclTensor
                std::vector<int16_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int16_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int16_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_FLOAT16) {
				// 创建self aclTensor
                std::vector<int16_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int16_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int16_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_DOUBLE) {
				// 创建self aclTensor
                std::vector<double> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<double> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<double> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_INT8) {
				// 创建self aclTensor
                std::vector<int8_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int8_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int8_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_INT16) {
				// 创建self aclTensor
                std::vector<int16_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int16_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int16_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_INT32) {
				// 创建self aclTensor
                std::vector<int32_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int32_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int32_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_INT64) {
				// 创建self aclTensor
                std::vector<int64_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int64_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int64_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_UINT8) {
				// 创建self aclTensor
                std::vector<uint8_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<uint8_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<uint8_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} else if (dtype0 == aclDataType::ACL_BOOL) {
				// 创建self aclTensor
                std::vector<int8_t> selfHostData (selfShape_size , 1);
                ret = CreateAclTensor_New(selfHostData , selfShape, &selfDeviceAddr, dtype0, format0, &self);
                cstSelfShapeSz = (uint64_t)GetShapeSize(selfShape);
                cstSelfDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建maxOut aclTensor
                std::vector<int8_t> maxOutHostData (maxOutShape_size , 1);
                ret = CreateAclTensor_New(maxOutHostData , maxOutShape, &maxOutDeviceAddr, dtype0, format0, &maxOut);
                cstMaxOutShapeSz = (uint64_t)GetShapeSize(maxOutShape);
                cstMaxOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
				// 创建minOut aclTensor
                std::vector<int8_t> minOutHostData (minOutShape_size , 1);
                ret = CreateAclTensor_New(minOutHostData , minOutShape, &minOutDeviceAddr, dtype0, format0, &minOut);
                cstMinOutShapeSz = (uint64_t)GetShapeSize(minOutShape);
                cstMinOutDtype = (int)(dtype0);
                CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
			} 

			//3. 调用aclnnAminmaxAll执行
            // 调用aclnnAminmaxAll第一段接口
                        // ===== [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);
            }
            cstMaxOutBytes = (maxOutDeviceAddr != nullptr) ? (size_t)cstMaxOutShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstMaxOutDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=maxOut, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstMaxOutDtype, ConsistentDtypeStr((aclDataType)cstMaxOutDtype), (unsigned long long)cstMaxOutShapeSz, cstMaxOutBytes, ConsistentSizeOfDtype((aclDataType)cstMaxOutDtype));
            if (cstMaxOutBytes == 0 && maxOutDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=maxOut addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstMaxOutSaveBuf = (cstMaxOutBytes > 0) ? malloc(cstMaxOutBytes) : nullptr;
            if (cstMaxOutSaveBuf != nullptr && maxOutDeviceAddr != nullptr) {
                aclrtMemcpy(cstMaxOutSaveBuf, cstMaxOutBytes, maxOutDeviceAddr, cstMaxOutBytes, ACL_MEMCPY_DEVICE_TO_HOST);
            }
            cstMinOutBytes = (minOutDeviceAddr != nullptr) ? (size_t)cstMinOutShapeSz * (size_t)ConsistentSizeOfDtype((aclDataType)cstMinOutDtype) : 0;
            printf("[FUZZ_CONSISTENT_BYTES] case_%d: tensor=minOut, dtype=%d(%s), shapeSize=%llu, cstBytes=%zu, ConsistentSizeOfDtype=%zu\n", fuzzi, (int)cstMinOutDtype, ConsistentDtypeStr((aclDataType)cstMinOutDtype), (unsigned long long)cstMinOutShapeSz, cstMinOutBytes, ConsistentSizeOfDtype((aclDataType)cstMinOutDtype));
            if (cstMinOutBytes == 0 && minOutDeviceAddr != nullptr) {
                printf("[FUZZ_CONSISTENT_BYTES] case_%d: WARN tensor=minOut addr!=nullptr but cstBytes=0, dtype may be uncovered by ConsistentSizeOfDtype\n", fuzzi);
            }
            cstMinOutSaveBuf = (cstMinOutBytes > 0) ? malloc(cstMinOutBytes) : nullptr;
            if (cstMinOutSaveBuf != nullptr && minOutDeviceAddr != nullptr) {
                aclrtMemcpy(cstMinOutSaveBuf, cstMinOutBytes, minOutDeviceAddr, cstMinOutBytes, 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 (cstMaxOutSaveBuf != nullptr && maxOutDeviceAddr != nullptr) {
                    aclrtMemcpy(maxOutDeviceAddr, cstMaxOutBytes, cstMaxOutSaveBuf, cstMaxOutBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                if (cstMinOutSaveBuf != nullptr && minOutDeviceAddr != nullptr) {
                    aclrtMemcpy(minOutDeviceAddr, cstMinOutBytes, cstMinOutSaveBuf, cstMinOutBytes, ACL_MEMCPY_HOST_TO_DEVICE);
                }
                ret = aclnnAminmaxAllGetWorkspaceSize(self, minOut, maxOut, &workspaceSize, &executor);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnAminmaxAll 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 = aclnnAminmaxAll(workspaceAddr, workspaceSize, executor, stream);
                if (ret != ACL_SUCCESS) {
                    printf("aclnnAminmaxAll 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 (maxOutDeviceAddr != nullptr && cstMaxOutBytes > 0) {
                    cstMaxOutHashArr[cstRun] = ConsistentHashDeviceTensor(maxOutDeviceAddr, cstMaxOutBytes);
                }
                if (minOutDeviceAddr != nullptr && cstMinOutBytes > 0) {
                    cstMinOutHashArr[cstRun] = ConsistentHashDeviceTensor(minOutDeviceAddr, cstMinOutBytes);
                }
                cstRunCount++;
                printf("aclnnAminmaxAll 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 cstMaxOutSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstMaxOutHashArr[cstR] != cstMaxOutHashArr[0]) { cstMaxOutSame = false; break; }
                }
                bool cstMinOutSame = true;
                for (int cstR = 1; cstR < CONSISTENT_RUN_TIMES; cstR++) {
                    if (cstMinOutHashArr[cstR] != cstMinOutHashArr[0]) { cstMinOutSame = false; break; }
                }
                bool cstAllSame = cstSelfSame && cstMaxOutSame && cstMinOutSame;
                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 (!cstMaxOutSame) printf("maxOut ");
                    if (!cstMinOutSame) printf("minOut ");
                    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: maxOut hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstMaxOutHashArr[cstR]);
                    printf("\n");
                    printf("[FUZZ_CONSISTENT_DETAIL] case_%d: minOut hash:", fuzzi);
                    for (int cstR = 0; cstR < CONSISTENT_RUN_TIMES; cstR++) printf(" run%d=%llu", cstR, (unsigned long long)cstMinOutHashArr[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(format0)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                    printf("maxOut: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstMaxOutDtype), ConsistentFormatStr(format0)); printf("shapeSize=%llu", (unsigned long long)cstMaxOutShapeSz); printf("; ");
                    printf("minOut: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstMinOutDtype), ConsistentFormatStr(format0)); printf("shapeSize=%llu", (unsigned long long)cstMinOutShapeSz); printf("; ");
                    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(format0)); printf("shapeSize=%llu", (unsigned long long)cstSelfShapeSz); printf("; ");
                printf("maxOut: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstMaxOutDtype), ConsistentFormatStr(format0)); printf("shapeSize=%llu", (unsigned long long)cstMaxOutShapeSz); printf("; ");
                printf("minOut: dtype=%s, format=%s, ", ConsistentDtypeStr((aclDataType)cstMinOutDtype), ConsistentFormatStr(format0)); printf("shapeSize=%llu", (unsigned long long)cstMinOutShapeSz); printf("; ");
                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 (cstMaxOutSaveBuf != nullptr) { free(cstMaxOutSaveBuf); cstMaxOutSaveBuf = nullptr; }
        if (cstMinOutSaveBuf != nullptr) { free(cstMinOutSaveBuf); cstMinOutSaveBuf = nullptr; }
aclDestroyTensor(self);
		aclDestroyTensor(maxOut);
		aclDestroyTensor(minOut);
		aclrtFree(selfDeviceAddr);
		aclrtFree(maxOutDeviceAddr);
		aclrtFree(minOutDeviceAddr);
		if (workspaceSize > 0) {
            aclrtFree(workspaceAddr);
        }
    }
    DT_FUZZ_END()
    WriteDTLog("aclnnAminmaxAll", "aclnnAminmaxAll 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("aclnnAminmaxAll", "aclnnAminmaxAll consistent op fuzz end \n");
    aclrtDestroyStream(stream);
    aclrtDestroyContext(context);
    aclrtResetDevice(deviceId);
    aclFinalize();   
    return 0;
}
extern "C" int aclnnAminmaxAll_consistent_fuzz(int argc, char *argv[]);
int aclnnAminmaxAll_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;
        cout << "The second arg stand for testcase num  \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("aclnnAminmaxAll : %d \n", aclnnAminmaxAll_consistent(fuzz_num));

    return 0;
}