aclnnMultiScaleDeformableAttentionGrad

Applicable Products

Product	Supported
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√

Function

Operator function:

The forward operator of MultiScaleDeformableAttention mainly traverses different sampling points of feature maps of different sizes through parameters such as sample locations, attention weights, mapped value features, the starting index positions of multi-scale features, and the spatial sizes of multi-scale feature maps (to convert sample locations from normalized values to absolute positions). The backward operator computes the gradients of the inputs based on the contribution of the forward inputs to the outputs and the initial gradients.

Formula:

Function Prototype

Each operator is divided into a two-phase API. You must first call the "aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call the "aclnnMultiScaleDeformableAttentionGrad" API to perform the computation.

aclnnStatus aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize(
    const aclTensor* value,
    const aclTensor* spatialShape,
    const aclTensor* levelStartIndex,
    const aclTensor* location,
    const aclTensor* attnWeight,
    const aclTensor* gradOutput,
    aclTensor*       gradValue,
    aclTensor*       gradLocation,
    aclTensor*       gradAttnWeight,
    uint64_t*        workspaceSize,
    aclOpExecutor**  executor)

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

aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize

Parameters:

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
value	Input	Feature values of the feature map.	-	FLOAT, FLOAT16, BFLOAT16	ND	(bs, num_keys, num_heads, channels) where bs is the batch size, num_keys is the size of the feature map, num_heads is the number of heads, and channels is the dimension of the feature map	√
spatialShape	Input	Stores the height and width of the feature map at each scale.	-	INT32, INT64	ND	(num_levels, 2) where num_levels is the count of feature maps, and 2 represents H and W respectively.	√
levelStartIndex	Input	Starting index of each feature map.	-	INT32, INT64	ND	(num_levels)	√
location	Input	Sampling point position tensor, which stores the coordinate position of each sampling point.	-	FLOAT, FLOAT16, BFLOAT16	ND	(bs, num_queries, num_heads, num_levels, num_points, 2) where num_queries is the number of queries, num_points is the number of sampling points, and 2 represents y and x respectively.	√
attnWeight	Input	Sampling point weight tensor.	-	FLOAT, FLOAT16, BFLOAT16	ND	(bs, num_queries, num_heads, num_levels, num_points)	√
gradOutput	Input	Gradient of the forward output, which is also the initial gradient of the backward operator.	-	FLOAT, FLOAT16, BFLOAT16	ND	(bs, num_queries, num_heads*channels)	√
gradValue	Output	Gradient corresponding to the input value.	The shape is consistent with value.	FLOAT, FLOAT16, BFLOAT16	ND	-	√
gradLocation	Output	Gradient corresponding to the input location.	Shape is the same as that of location.	FLOAT, FLOAT16, BFLOAT16	ND	-	√
gradAttnWeight	Output	Gradient corresponding to the input attnWeight.	The shape is the same as that of attnWeight.	FLOAT, FLOAT16, BFLOAT16	ND	-	√
workspaceSize	Output	Returns the workspace size to be requested on the Device side.	-	-	-	-	-
executor	Output	Returns the op executor, which contains the operator computation flow.	-	-	-	-	-

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 Value	Error Code	Description
ACLNN_ERR_PARAM_NULLPTR	161001	The input or output passed in is a null pointer.
ACLNN_ERR_PARAM_INVALID	161002	The data types of the input and output are not within the supported range.
		The data types of the input and output are inconsistent.
		The constraints specified in the Constraints section are not met.

aclnnMultiScaleDeformableAttentionGrad

Parameters:

Parameter	Input/Output	Description
workspace	Input	Address of the workspace memory applied for on the Device side.
workspaceSize	Input	Size of the workspace applied for on the Device side, obtained through the first-phase API aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize.
executor	Input	Operator executor, which contains the operator computation flow.
stream	Input	Specifies the stream for task execution.

Return value:

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

Constraints

Deterministic computation
- aclnnMultiScaleDeformableAttentionGrad defaults to a non-deterministic implementation and does not support deterministic computation currently.
The number of channels satisfies channels % 8 = 0 and channels <= 256.
The number of queries satisfies num_queries < 500000.
The number of feature maps, num_levels, <= 16.
The number of heads, num_heads, <= 16.
The number of sampling points, num_points, <= 16.

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_multi_scale_deformable_attention_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 shapeSize = 1;
    for (auto i : shape) {
        shapeSize *= i;
  }
    return shapeSize;
}

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 to the device using aclrtMemcpy.
    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 aclCreateTensor 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 API list.
    // Set deviceId based on the actual device.
    int32_t deviceId = 0;
    aclrtStream stream;
    auto ret = Init(deviceId, &stream);
    // Perform check as needed.
    CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // 2. Construct the input and output. Customize the construction based on the API interface.

    std::vector<int64_t> valueShape = {1, 1, 1, 8};
    std::vector<int64_t> spatialShapeShape = {1, 2};
    std::vector<int64_t> levelStartIndexShape = {1};
    std::vector<int64_t> locationShape = {1, 32, 1, 1, 1, 2};
    std::vector<int64_t> attnWeightShape = {1, 32, 1, 1, 1};
    std::vector<int64_t> gradOutputShape = {1, 32, 8};
    std::vector<int64_t> gradValueShape = {1, 1, 1, 8};
    std::vector<int64_t> gradLocationShape = {1, 32, 1, 1, 1, 2};
    std::vector<int64_t> gradAttnWeightShape = {1, 32, 1, 1, 1};
    void* valueDeviceAddr = nullptr;
    void* spatialShapeDeviceAddr = nullptr;
    void* levelStartIndexDeviceAddr = nullptr;
    void* locationDeviceAddr = nullptr;
    void* attnWeightDeviceAddr = nullptr;
    void* gradOutputDeviceAddr = nullptr;
    void* gradValueDeviceAddr = nullptr;
    void* gradLocationDeviceAddr = nullptr;
    void* gradAttnWeightDeviceAddr = nullptr;
    aclTensor* value = nullptr;
    aclTensor* spatialShape = nullptr;
    aclTensor* levelStartIndex = nullptr;
    aclTensor* location = nullptr;
    aclTensor* attnWeight = nullptr;
    aclTensor* gradOutput = nullptr;
    aclTensor* gradValue = nullptr;
    aclTensor* gradLocation = nullptr;
    aclTensor* gradAttnWeight = nullptr;
    std::vector<float> valueHostData = {1, 1, 1, 1, 1, 1, 1, 1};
    std::vector<float> spatialShapeHostData = {1, 1};
    std::vector<float> levelStartIndexHostData = {0};
    std::vector<float> gradValueHostData = {0, 0, 0, 0, 0, 0, 0, 0};
    std::vector<float> gradLocationHostData(GetShapeSize(gradLocationShape), 0);
    std::vector<float> gradAttnWeightHostData(GetShapeSize(gradAttnWeightShape), 0);
    std::vector<float> locationHostData(GetShapeSize(locationShape), 0);
    std::vector<float> attnWeightHostData(GetShapeSize(attnWeightShape), 1);
    std::vector<float> gradOutputHostData(GetShapeSize(gradOutputShape), 1);

    // value aclTensor
    ret = CreateAclTensor(valueHostData, valueShape, &valueDeviceAddr, aclDataType::ACL_FLOAT, &value);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the spatialShape aclTensor.
    ret = CreateAclTensor(spatialShapeHostData, spatialShapeShape, &spatialShapeDeviceAddr, aclDataType::ACL_INT32, &spatialShape);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the levelStartIndex aclTensor.
    ret = CreateAclTensor(levelStartIndexHostData, levelStartIndexShape, &levelStartIndexDeviceAddr, aclDataType::ACL_INT32, &levelStartIndex);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the location aclTensor.
    ret = CreateAclTensor(locationHostData, locationShape, &locationDeviceAddr, aclDataType::ACL_FLOAT, &location);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the attnWeight aclTensor.
    ret = CreateAclTensor(attnWeightHostData, attnWeightShape, &attnWeightDeviceAddr, aclDataType::ACL_FLOAT, &attnWeight);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // gradOutput aclTensor
    ret = CreateAclTensor(gradOutputHostData, gradOutputShape, &gradOutputDeviceAddr, aclDataType::ACL_FLOAT, &gradOutput);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the gradValue aclTensor.
    ret = CreateAclTensor(gradValueHostData, gradValueShape, &gradValueDeviceAddr, aclDataType::ACL_FLOAT, &gradValue);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the gradLocation aclTensor.
    ret = CreateAclTensor(gradLocationHostData, gradLocationShape, &gradLocationDeviceAddr, aclDataType::ACL_FLOAT, &gradLocation);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the gradAttnWeight aclTensor.
    ret = CreateAclTensor(gradAttnWeightHostData, gradAttnWeightShape, &gradAttnWeightDeviceAddr, aclDataType::ACL_FLOAT, &gradAttnWeight);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    // 3. Call the CANN operator library API. Replace it with the specific API.
    uint64_t workspaceSize = 0;
    aclOpExecutor* executor;
    // Call the first-phase API of aclnnMultiScaleDeformableAttentionGrad.
    ret = aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize(value, spatialShape, levelStartIndex, location, attnWeight, gradOutput, gradValue, gradLocation, gradAttnWeight, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMultiScaleDeformableAttentionGradGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
    // Allocate device memory based on the workspaceSize computed 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 aclnnMultiScaleDeformableAttentionGrad.
    ret = aclnnMultiScaleDeformableAttentionGrad(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMultiScaleDeformableAttentionGrad failed. ERROR: %d\n", ret); return ret);
    // 4. (Boilerplate) Synchronize and 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 device memory to host memory. Modify based on the specific API definition.
    auto size = GetShapeSize(gradValueShape);
    std::vector<float> resultData(size, 0);
    ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), gradValueDeviceAddr, size * 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; i++) {
        LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
    }

    // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
    aclDestroyTensor(value);
    aclDestroyTensor(spatialShape);
    aclDestroyTensor(levelStartIndex);
    aclDestroyTensor(location);
    aclDestroyTensor(attnWeight);
    aclDestroyTensor(gradOutput);
    aclDestroyTensor(gradValue);
    aclDestroyTensor(gradLocation);
    aclDestroyTensor(gradAttnWeight);

    // 7. Release device resources. Modify based on the specific API definition.
    aclrtFree(valueDeviceAddr);
    aclrtFree(spatialShapeDeviceAddr);
    aclrtFree(levelStartIndexDeviceAddr);
    aclrtFree(locationDeviceAddr);
    aclrtFree(attnWeightDeviceAddr);
    aclrtFree(gradOutputDeviceAddr);
    aclrtFree(gradValueDeviceAddr);
    aclrtFree(gradLocationDeviceAddr);
    aclrtFree(gradAttnWeightDeviceAddr);
    if (workspaceSize > 0) {
        aclrtFree(workspaceAddr);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}