aclnnRmsNormGrad

Applicable Products

Product	Supported
Ascend 950PR/Ascend 950DT	√
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	×
Atlas inference products	√
Atlas training products	×

Function

Description: Backward computation of aclnnRmsNorm. Computes the gradient of RmsNorm, i.e., computes the gradient of the input tensor during backward propagation.
Operator formula:
- Forward formula:
$\operatorname{RmsNorm}(x_i)=\frac{1}{\operatorname{Rms}(\mathbf{x})} * x_i * g_i, \quad \text { where } \operatorname{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+eps}$
- Backward derivation:
$dx_i= (dy_i * g_i - \frac{1}{\operatorname{Rms}(\mathbf{x})} * x_i * \operatorname{Mean}(\mathbf{y})) * \frac{1} {\operatorname{Rms}(\mathbf{x})}, \quad \text { where } \operatorname{Mean}(\mathbf{y}) = \frac{1}{n}\sum_{i=1}^n (dy_i * g_i * x_i * \frac{1}{\operatorname{Rms}(\mathbf{x})})$ $dg_i = \frac{1}{\operatorname{Rms}(\mathbf{x})} * x_i * dy_i$

Function Prototype

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

aclnnStatus aclnnRmsNormGradGetWorkspaceSize(
  const aclTensor *dy,
  const aclTensor *x,
  const aclTensor *rstd,
  const aclTensor *gamma,
  const aclTensor *dxOut,
  const aclTensor *dgammaOut,
  uint64_t        *workspaceSize,
  aclOpExecutor  **executor)

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

aclnnRmsNormGradGetWorkspaceSize

Parameters

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (shape)	Non-contiguous Tensor
dy (aclTensor*)	Input	Indicates the gradient returned from backward propagation, which corresponds to `dy` in the formula.	Supports null Tensor.	FLOAT32, FLOAT16, BFLOAT16	ND	1-8	√
x (aclTensor*)	Input	Indicates the input of the forward operator, representing the data to be normalized, corresponding to `x` in the formula.	Supports null Tensor. The shape remains consistent with the shape of input parameter `dy`.	FLOAT32, FLOAT16, BFLOAT16	ND	1-8	√
rstd (aclTensor*)	Input	Indicates the intermediate computation result of the forward operator, corresponding to `Rms(x)` in the formula.	Supports null tensor. The shape must satisfy rstd_shape = x_shape[0:n], where n < x_shape.dims() and n remains consistent with that of `gamma`.	FLOAT32	ND	1-8	√
gamma (aclTensor*)	Input	Scaling factor (weight) used by the forward operator for normalization computation, corresponding to `gamma` in the formula.	Supports null Tensor. The shape must satisfy gamma_shape = x_shape[n:], n < x_shape.dims().	FLOAT32, FLOAT16, BFLOAT16	ND	1-8	√
dxOut (aclTensor*)	Output	Indicates the gradient of input `x`, corresponding to `dx` in the formula.	Supports null tensor. The shape remains consistent with input parameter `dy`.	FLOAT32, FLOAT16, BFLOAT16	ND	1-8	√
dgammaOut (aclTensor*)	Output	Indicates the gradient of `gamma`, corresponding to `dg` in the formula.	Supports null tensor. The shape remains consistent with the shape of input parameter `gamma`.	FLOAT32	ND	1-8	√
workspaceSize (uint64_t*)	Output	Returns the workspace size to be applied on the Device side.	-	-	-	-	-
executor (aclOpExecutor**)	Output	Returns the op executor, which contains the operator computation flow.	-	-	-	-	-

Atlas inference products: The data type of parameters dy, x, gamma, and dxOut does not support BFLOAT16.

Return Value

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

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

Return Code	Error Code	Description
ACLNN_ERR_PARAM_NULLPTR	161001	If the input parameter is a required input, output, or required attribute and is a null pointer, 161001 is returned.
ACLNN_ERR_PARAM_INVALID	161002	The data type of the input or output is not within the supported range.
ACLNN_ERR_INNER_TILING_ERROR	561002	The parameter does not meet the requirements in Parameters.

aclnnRmsNormGrad

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 aclnnRmsNormGradGetWorkspaceSize.
executor	Input	Op executor that contains the operator computation flow.
stream	Input	Stream that executes the task.

Return Value

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

Constraints

Atlas inference products: The last axis length of the x, dy, and gamma inputs must be greater than or equal to 32 bytes.

Supported data types for each product:

Atlas A2 training products/Atlas A2 inference products, Atlas A3 training products/Atlas A3 inference products, Ascend 950PR/Ascend 950DT:

`dy` data type	`x` data type	`rstd` data type	`gamma` data type	`dxOut` data type	`dgammaOut` data type
FLOAT16	FLOAT16	FLOAT32	FLOAT32	FLOAT16	FLOAT32
BFLOAT16	BFLOAT16	FLOAT32	FLOAT32	BFLOAT16	FLOAT32
FLOAT16	FLOAT16	FLOAT32	FLOAT16	FLOAT16	FLOAT32
FLOAT32	FLOAT32	FLOAT32	FLOAT32	FLOAT32	FLOAT32
BFLOAT16	BFLOAT16	FLOAT32	BFLOAT16	BFLOAT16	FLOAT32

Atlas inference products:

dy Data Type x Data Type rstd Data Type gamma Data Type dxOut Data Type dgammaOut Data Type
FLOAT16 FLOAT16 FLOAT32 FLOAT16 FLOAT16 FLOAT32
FLOAT32 FLOAT32 FLOAT32 FLOAT32 FLOAT32 FLOAT32

Deterministic computation:
- aclnnRmsNormGrad defaults to a non-deterministic implementation. Deterministic computation can be enabled by calling aclrtCtxSetSysParamOpt.

`dy` Data Type	`x` Data Type	`rstd` Data Type	`gamma` Data Type	`dxOut` Data Type	`dgammaOut` Data Type
FLOAT16	FLOAT16	FLOAT32	FLOAT16	FLOAT16	FLOAT32
FLOAT32	FLOAT32	FLOAT32	FLOAT32	FLOAT32	FLOAT32

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_rms_norm_grad.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(const std::vector<int64_t>& shape)
{
    int64_t shape_size = 1;
    for (auto i : shape) {
        shape_size *= i;
    }
    return shape_size;
}

int Init(int32_t deviceId, aclrtStream* stream)
{
    // Boilerplate: resource 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(
    const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr, aclDataType dataType,
    aclTensor** tensor)
{
    auto size = GetShapeSize(shape) * sizeof(T);
    // Call aclrtMalloc to apply for 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);

    // Copy data from the host side to the device 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);

    // Compute strides for 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_NCDHW, shape.data(),
        shape.size(), *deviceAddr);
    return 0;
}

int main()
{
    // 1. (Boilerplate) Initialize the device/stream. See the ACL API manual.
    // Set deviceId based on your actual device.
    int32_t deviceId = 0;
    aclrtStream stream;
    auto ret = Init(deviceId, &stream);
    // Handle the check based on your needs.
    CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // 2. Construct the input and output. Custom construction is required based on the API interface.
    std::vector<int64_t> gradInputShape = {2, 1, 16};
    std::vector<int64_t> xInputShape = {2, 1, 16};
    std::vector<int64_t> rstdInputShape = {2};
    std::vector<int64_t> gammaInputShape = {16};
    std::vector<int64_t> dxOutputShape = {2, 1, 16};
    std::vector<int64_t> dgammaOutputShape = {16};

    void* gradInputDeviceAddr = nullptr;
    void* xInputDeviceAddr = nullptr;
    void* rstdInputDeviceAddr = nullptr;
    void* gammaInputDeviceAddr = nullptr;
    void* dxOutDeviceAddr = nullptr;
    void* dgammaOutDeviceAddr = nullptr;

    aclTensor* gradInput = nullptr;
    aclTensor* xInput = nullptr;
    aclTensor* rstdInput = nullptr;
    aclTensor* gammaInput = nullptr;
    aclTensor* dxOut = nullptr;
    aclTensor* dgammaOut = nullptr;

    std::vector<float> gradInputHostData = {1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15, 16,
                                            17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32};
    std::vector<float> xInputHostData = {1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15, 16,
                                         17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32};
    std::vector<float> rstdInputHostData = {1, 2};
    std::vector<float> gammaInputHostData = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};

    std::vector<float> dxOutHostData = {1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15, 16,
                                        17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32};
    std::vector<float> dgammaOutHostData = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};

    std::vector<int64_t> output1SizeData = {2, 1, 16};
    std::vector<int64_t> output2SizeData = {16};
    std::vector<int64_t> input1SizeData = {2, 1, 16};
    std::vector<int64_t> input2SizeData = {2};
    std::vector<int64_t> input3SizeData = {16};

    ret = CreateAclTensor(gradInputHostData, input1SizeData, &gradInputDeviceAddr, aclDataType::ACL_FLOAT, &gradInput);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    ret = CreateAclTensor(xInputHostData, input1SizeData, &xInputDeviceAddr, aclDataType::ACL_FLOAT, &xInput);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    ret = CreateAclTensor(rstdInputHostData, input2SizeData, &rstdInputDeviceAddr, aclDataType::ACL_FLOAT, &rstdInput);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    ret =
        CreateAclTensor(gammaInputHostData, input3SizeData, &gammaInputDeviceAddr, aclDataType::ACL_FLOAT, &gammaInput);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    ret = CreateAclTensor(dxOutHostData, output1SizeData, &dxOutDeviceAddr, aclDataType::ACL_FLOAT, &dxOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    ret = CreateAclTensor(dgammaOutHostData, output2SizeData, &dgammaOutDeviceAddr, aclDataType::ACL_FLOAT, &dgammaOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    // 3. Call the CANN operator library API. Modify it to the specific API.
    uint64_t workspaceSize = 0;
    aclOpExecutor* executor;
    // Call the first-phase API of aclnnRmsNormGrad.
    ret = aclnnRmsNormGradGetWorkspaceSize(
        gradInput, xInput, rstdInput, gammaInput, dxOut, dgammaOut, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnRmsNormGradGetWorkspaceSize 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 aclnnRmsNormGrad.
    ret = aclnnRmsNormGrad(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnRmsNormGrad failed. ERROR: %d\n", ret); return ret);
    // 4. (Boilerplate) Wait for task execution to complete synchronously.
    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-side memory to the host side. Modify based on the specific API definition.
    auto size_dx = GetShapeSize(gradInputShape);
    std::vector<float> resultData1(size_dx, 0);
    ret = aclrtMemcpy(
        resultData1.data(), resultData1.size() * sizeof(resultData1[0]), dxOutDeviceAddr, size_dx * sizeof(float),
        ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < size_dx; i++) {
        LOG_PRINT("result[%ld] is: %f\n", i, resultData1[i]);
    }
    auto size_dgamma = GetShapeSize(gammaInputShape);
    std::vector<float> resultData2(size_dgamma, 1);
    ret = aclrtMemcpy(
        resultData2.data(), resultData2.size() * sizeof(resultData2[0]), dgammaOutDeviceAddr,
        size_dgamma * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < size_dgamma; i++) {
        LOG_PRINT("result[%ld] is: %f\n", i, resultData2[i]);
    }

    // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
    aclDestroyTensor(gradInput);
    aclDestroyTensor(xInput);
    aclDestroyTensor(rstdInput);
    aclDestroyTensor(gammaInput);
    aclDestroyTensor(dxOut);
    aclDestroyTensor(dgammaOut);

    // 7. Release device resources. Modify based on the specific API definition.
    aclrtFree(gradInputDeviceAddr);
    aclrtFree(xInputDeviceAddr);
    aclrtFree(rstdInputDeviceAddr);
    aclrtFree(gammaInputDeviceAddr);
    aclrtFree(dxOutDeviceAddr);
    aclrtFree(dgammaOutDeviceAddr);

    if (workspaceSize > 0) {
        aclrtFree(workspaceAddr);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}