aclnnSilentCheckV2

Applicable Products

Function

• Operator function:

  Added the aclnnSilentCheckV2 API. This version changes the computation logic compared with the aclnnSilentCheck API, and compares val with the newly added Markov inequality threshold based on the stepRef parameter for verification.

• Formula:

  • If stepRef==0:

    • If the current input val is inf/-inf/nan, or val exceeds avgRef * cThreshL1, it is identified as an L1 Level Fault, and an ERROR log is printed. If the environment variable npuAsdDetect is 1, avgRef and stepRef are updated and the function returns normally; otherwise, inputGradRef is set to zero and checkpoint-based training recovery is triggered.

    • If the current input val exceeds avgRef * cThreshL2, it is identified as an L2 Level Alarm; a WARNING log is printed, and avgRef and stepRef are updated before returning normally.

  • If stepRef > 0:

    • If the current input val is inf/-inf/nan, or val exceeds the Markov inequality threshold (avgRef/(1-beta1)^stepRef) * cThreshL1, it is identified as an L1 Level Fault, and an ERROR log is printed. If the environment variable npuAsdDetect is 2, checkpoint-based training recovery is triggered; if the environment variable npuAsdDetect is 1, avgRef and stepRef are updated and the function returns normally.

    • If the current input val exceeds the Markov inequality threshold (avgRef/(1-beta1)^stepRef) * cThreshL2, it is identified as an L2 Level Alarm; a WARNING log is printed, and avgRef and stepRef are updated before returning normally.

    • If no L1 Level Fault or L2 Level Alarm is triggered, under normal conditions: if npuAsdDetect is 3, the characteristic values of val are printed; otherwise, avgRef and stepRef are updated and the function returns normally.

  • stepRef indicates the detection count.

Function Prototype

Each operator is divided into a two-phase API. The "aclnnSilentCheckV2GetWorkspaceSize" API must be called first to obtain the workspace size required for computation and the executor that encapsulates the operator computation flow. Then, the "aclnnSilentCheckV2" API is called to perform the computation.

aclnnStatus aclnnSilentCheckV2GetWorkspaceSize(
    const aclTensor *val,
    const aclTensor *max,
    aclTensor *avgRef,
    aclTensor *inputGradRef,
    aclTensor *stepRef,
    aclIntArray *dstSize,
    aclIntArray *dstStride,
    aclIntArray *dstOffset,
    float cThreshL1,
    float cThreshL2,
    float beta1,
    int32_t npuAsdDetect,
    aclTensor* result,
    uint64_t *workspaceSize,
    aclOpExecutor **executor)

aclnnStatus aclnnSilentCheckV2(
    void *workspace,
    uint64_t workspaceSize,
    aclOpExecutor *executor,
    aclrtStream stream)

aclnnSilentCheckV2GetWorkspaceSize

• Parameters:

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
  val	Input	The current input value.	-	FLOAT, FLOAT16, BFLOAT16	ND	1	√
  max	Input.	-	-	FLOAT, FLOAT16, BFLOAT16	ND	1	√
  avgRef	Input	-	-	FLOAT, FLOAT16, BFLOAT16	ND	1	√
  inputGradRef	Input	Gradient tensor of the model input.	-	FLOAT, FLOAT16, BFLOAT16	ND	-	√
  stepRef	Input	Current step count.	Negative numbers are not supported.	INT64	ND	Must be [1].	√
  dstSize	Input	-	-	INT64	ND	-	-
  dstStride	Input	-	-	INT64	ND	-	-
  dstOffset	Input	-	-	INT64	ND	-	-
  cThreshL1	Input	Absolute value threshold for triggering an L1 Level Fault.	Currently recommended value: 1000000.	FLOAT	-	-	-
  cThreshL2	Input	Absolute numeric threshold for triggering an L2 level fault.	Currently, the recommended value is 10000, and cThreshL1 > cThreshL2.	FLOAT	-	-	-
  beta1	Input	-	Currently recommended value is 0.99, with a range of 0 < beta1 < 1. It is recommended that the value be infinitely close to 1.	FLOAT	-	-	-
  npuAsdDetect	Input	Environment variable.	-	INT32	-	-	-
  result	Output	Returns the silent check result.	-	INT32	ND	-	√
  workspaceSize	Output	Returns the workspace size to be allocated on the Device side.	-	-	-	-	-
  executor	Output	Returns the operator executor, which contains the operator computation flow.	-	-	-	-	-

• Return value:

  aclnnStatus: return code. For details, see aclnn Return Codes.

  The first-phase API performs input parameter validation and returns an error in the following scenarios:

  Return Value	Error Code	Description
  ACLNN_ERR_PARAM_NULLPTR	161001	The passed val, max, avgRef, inputGradRef, stepRef, dstSize, dstStride, and dstOffset are null pointers.
  ACLNN_ERR_PARAM_INVALID	161002	The data type of val, max, avgRef, inputGradRef, stepRef, dstSize, dstStride, or dstOffset is not within the supported range.
  		The shape of val, max, avgRef, inputGradRef, stepRef, dstSize, dstStride, or dstOffset does not meet the requirement.
  		The value of cThreshL1, cThreshL2, or beta1 is not within the supported range.

aclnnSilentCheck

• Parameters:

  Parameter	Input/Output	Description
  workspace	Input	Address of the workspace memory applied on the Device side.
  workspaceSize	Input	Size of the workspace applied on the Device side, obtained from the first-phase API aclnnSilentCheckGetWorkspaceSize.
  executor	Input	Operator executor that contains the operator computation flow.
  stream	Input	Specifies the stream on which the task is executed.

• Return value:

  aclnnStatus: return code. For details, see aclnn Return Codes.

Constraints

• Deterministic computation:

  • aclnnSilentCheckV2 defaults to a deterministic implementation.

Example

The following provides example code for reference only. For details about compilation and running, see Compile and Run Samples.

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_silent_check_v2.h"

#define CHECK_RET(cond, return_expr) \
    do {                               \
        if (!(cond)) {                   \
        return_expr;                   \
        }                                \
    } while (0)

#define LOG_PRINT(message, ...)     \
    do {                              \
        printf(message, ##__VA_ARGS__); \
    } while (0)

int64_t GetShapeSize(std::vector<int64_t>& shape) {
    int64_t shapeSize = 1;
    for (auto i : shape) {
        shapeSize *= i;
  }
    return shapeSize;
}

void PrintOutResult(std::vector<int64_t> &shape, void** deviceAddr) {
  auto size = GetShapeSize(shape);
  std::vector<int32_t> resultData(size, 0);
  auto ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]),
                         *deviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return);
  for (int64_t i = 0; i < size; i++) {
    LOG_PRINT("result[%ld] is: %d\n", i, resultData[i]);
  }
}

int Init(int32_t deviceId, aclrtStream* stream) {
    // Boilerplate for ACL initialization.
    auto ret = aclInit(nullptr);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
    ret = aclrtSetDevice(deviceId);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
    ret = aclrtCreateStream(stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
    return 0;
}

template <typename T>
int CreateAclTensor(std::vector<T>& hostData, std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) {
    auto size = GetShapeSize(shape) * sizeof(T);
    // Call aclrtMalloc to allocate device-side memory.
    auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

    // Call aclrtMemcpy to copy host-side data to device-side memory.
    ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

    // Calculate the strides of a contiguous tensor.
    std::vector<int64_t> strides(shape.size(), 1);
    for (int64_t i = shape.size() - 2; i >= 0; i--) {
       strides[i] = shape[i + 1] * strides[i + 1];
    }

    // Call the aclCreateTensor API to create an aclTensor.
    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                              shape.data(), shape.size(), *deviceAddr);
    return 0;
}

int main() {
    // 1. (Boilerplate) Initialize the device/stream. For details, see the acl external API list.
    // Set deviceId based on the actual device.
    int32_t deviceId = 0;
    aclrtStream stream;
    auto ret = Init(deviceId, &stream);
    // Handle the check based on your requirements.
    CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // 2. Construct inputs and outputs. Customize the construction based on the API.
    std::vector<int64_t> valShape = {1};
    std::vector<int64_t> maxShape = {1};
    std::vector<int64_t> avgRefShape = {1};
    std::vector<int64_t> inputGradRefShape = {2, 3};
    std::vector<int64_t> stepRefShape = {1};
    std::vector<int64_t> resultShape = {1};
    void* valDeviceAddr = nullptr;
    void* maxDeviceAddr = nullptr;
    void* avgRefDeviceAddr = nullptr;
    void* inputGradRefDeviceAddr = nullptr;
    void* stepRefDeviceAddr = nullptr;
    void* resultDeviceAddr = nullptr;
    aclTensor* val = nullptr;
    aclTensor* max = nullptr;
    aclTensor* avgRef = nullptr;
    aclTensor* inputGradRef = nullptr;
    aclTensor* stepRef = nullptr;
    aclTensor* result = nullptr;
    std::vector<float> valHostData = {160.0};
    std::vector<float> maxHostData = {400.0};
    std::vector<float> avgRefHostData = {200.0};
    std::vector<float> inputGradRefHostData = {0, 1, 2, 3, 4, 5};
    std::vector<int64_t> stepRefHostData = {0};
    std::vector<int64_t> dstSizeData = {2, 3};
    std::vector<int64_t> dstStrideData = {3, 1};
    std::vector<int64_t> dstOffsetData = {0};
    std::vector<int32_t> resultHostData = {0};
    aclIntArray* dstSize = aclCreateIntArray(dstSizeData.data(), 2);
    aclIntArray* dstStride = aclCreateIntArray(dstStrideData.data(), 2);
    aclIntArray* dstOffset = aclCreateIntArray(dstOffsetData.data(), 1);
    float cThreshL1 = 1000000;
    float cThreshL2 = 10000;
    float beta1 = 0.99;
    int32_t npuAsdDetect = 3;

    // Create the val aclTensor.
    ret = CreateAclTensor(valHostData, valShape, &valDeviceAddr, aclDataType::ACL_FLOAT, &val);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the max aclTensor.
    ret = CreateAclTensor(maxHostData, maxShape, &maxDeviceAddr, aclDataType::ACL_FLOAT, &max);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the avgRef aclTensor.
    ret = CreateAclTensor(avgRefHostData, avgRefShape, &avgRefDeviceAddr, aclDataType::ACL_FLOAT, &avgRef);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the inputGradRef aclTensor.
    ret = CreateAclTensor(inputGradRefHostData, inputGradRefShape, &inputGradRefDeviceAddr, aclDataType::ACL_FLOAT, &inputGradRef);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the stepRef aclTensor.
    ret = CreateAclTensor(stepRefHostData, stepRefShape, &stepRefDeviceAddr, aclDataType::ACL_INT64, &stepRef);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the result aclTensor.
    ret = CreateAclTensor(resultHostData, resultShape, &resultDeviceAddr, aclDataType::ACL_INT32, &result);
    if (result == nullptr) {
        std::cout << "result is nullptr!" << std::endl;
        return 0;
    }
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    // 3. Call the CANN operator library API. Replace with the specific HostApi.
    uint64_t workspaceSize = 0;
    aclOpExecutor* executor;
    // Call the first-phase API of aclnnSilentCheckV2.
    ret = aclnnSilentCheckV2GetWorkspaceSize(val, max, avgRef, inputGradRef, stepRef, dstSize, dstStride, dstOffset, cThreshL1, cThreshL2, beta1, npuAsdDetect, result, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSilentCheckV2GetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
    // Apply for device memory based on the workspaceSize calculated by the first-phase API.
    void* workspaceAddr = nullptr;
    if (workspaceSize > 0) {
        ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
        CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
    }
    // Call the second-phase API of aclnnSilentCheckV2.
    ret = aclnnSilentCheckV2(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSilentCheckV2 failed. ERROR: %d\n", ret); return ret);
    // 4. (Boilerplate) Synchronously wait for task execution to complete.
    ret = aclrtSynchronizeStream(stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
    // 5. Obtain the output value and copy the result from the device memory to the host memory. Modify based on the specific API definition.
    PrintOutResult(resultShape, &resultDeviceAddr);

    // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
    aclDestroyTensor(val);
    aclDestroyTensor(max);
    aclDestroyTensor(avgRef);
    aclDestroyTensor(inputGradRef);
    aclDestroyTensor(stepRef);
    aclDestroyIntArray(dstSize);
    aclDestroyIntArray(dstStride);
    aclDestroyIntArray(dstOffset);
    aclDestroyTensor(result);

    // 7. Release device resources. Modify based on the specific API definition.
    aclrtFree(valDeviceAddr);
    aclrtFree(maxDeviceAddr);
    aclrtFree(avgRefDeviceAddr);
    aclrtFree(inputGradRefDeviceAddr);
    aclrtFree(stepRefDeviceAddr);
    aclrtFree(resultDeviceAddr);
    if (workspaceSize > 0) {
        aclrtFree(workspaceAddr);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}