CANN/ops-nn快速层归一化算子

aclnnFastLayerNorm【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn 查看源码产品支持情况产品是否支持Ascend 950PR/Ascend 950DT×Atlas A3 训练系列产品/Atlas A3 推理系列产品√Atlas A2 训练系列产品/Atlas A2 推理系列产品√Atlas 200I/500 A2 推理产品×Atlas 推理系列产品×Atlas 训练系列产品×功能说明接口功能对指定层进行均值为0、标准差为1的归一化计算。aclnnFastLayerNorm接口相比aclnnLayerNorm接口整体性能提升了50%内存与GPU保持一致累加序优化导致精度存在差异。计算公式$$ out \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] eps}} * weightOptional biasOptional $$$$ meanOutOptional \mathrm{E}[x] $$$$ rstdOutOptional \frac{1}{ \sqrt{\mathrm{Var}[x] eps}} $$其中E[x]表示输入的均值Var[x]表示输入的方差。函数原型每个算子分为两段式接口必须先调用“aclnnFastLayerNormGetWorkspaceSize”接口获取计算所需workspace大小以及包含了算子计算流程的执行器再调用“aclnnFastLayerNorm”接口执行计算。aclnnStatus aclnnFastLayerNormGetWorkspaceSize( const aclTensor *input, const aclIntArray *normalizedShape, const aclTensor *weightOptional, const aclTensor *biasOptional, double eps, aclTensor *out, aclTensor *meanOutOptional, aclTensor *rstdOutOptional, uint64_t *workspaceSize, aclOpExecutor **executor)aclnnStatus aclnnFastLayerNorm( void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)aclnnFastLayerNormGetWorkspaceSize参数说明参数名输入/输出描述使用说明数据类型数据格式维度(shape)非连续TensorinputaclTensor*输入表示进行归一化计算的输入对应公式中的x。支持空Tensor。shape为[A1,...,Ai,R1,...,Rj]其中A1至Ai表示无需norm的维度R1至Rj表示需norm的维度。FLOAT32、FLOAT16、BFLOAT16ND1-8√normalizedShapeaclIntArray*输入表示需要进行norm计算的维度。值为[R1,...,Rj]长度小于等于输入input的shape长度不支持为空。且R1*R2*...*Rj小于等于583705600。INT64---weightOptionalaclTensor*输入可选输入参数表示进行归一化计算的权重。对应公式中的weightOptional。支持空Tensor。当weightOptional非空时数据类型与输入input一致或为FLOAT类型且当biasOptional存在时weightOptional与biasOptional的数据类型相同。shape与normalizedShape相等为[R1,...,Rj]。当weightOptional为空时接口内部会构造一个shape为[R1,...,Rj]数据全为1的Tensor。当biasOptional存在时weightOptional与biasOptional的数据类型相同。当biasOptional不存在时weightOptional与输入input的数据类型相同。FLOAT32、FLOAT16、BFLOAT16ND1-8√biasOptionalaclTensor*输入可选输入参数表示进行归一化计算的偏移量。对应公式中的biasOptional。支持空Tensor。当biasOptional非空时数据类型与输入input一致或为FLOAT类型且当weightOptional存在时biasOptional与weightOptional的数据类型相同。shape与normalizedShape相等为[R1,...,Rj]。当biasOptional为空时接口内部会构造一个shape为[R1,...,Rj]数据全为0的Tensor。当weightOptional存在时biasOptional与weightOptional的数据类型相同。当weightOptional不存在时biasOptional与输入input的数据类型相同。FLOAT32、FLOAT16、BFLOAT16ND1-8√epsdouble输入表示添加到分母中的值以确保数值稳定。对应公式中的eps。-----outaclTensor*输出表示进行归一化计算的结果。对应公式中的out。支持空Tensor。数据类型与input的数据类型保持一致。shape需要与input的shape相等为[A1,...,Ai,R1,...,Rj]。FLOAT32、FLOAT16、BFLOAT16ND1-8√meanOutOptionalaclTensor*输出可选输出表示进行归一化后的均值。对应公式中的meanOutOptional。支持空Tensor。数据类型与input的数据类型保持一致。当rstdOutOptional存在时与rstdOutOptional的shape相同shape为[A1,...,Ai,1,...,1]Ai后共有j个1与需要norm的轴长度保持相同。FLOAT32、FLOAT16、BFLOAT16ND1-8√rstdOutOptionalaclTensor*输出可选输出表示进行归一化后的标准差倒数。对应公式中的rstdOutOptional。支持空Tensor。数据类型与input的数据类型保持一致。当meanOutOptional存在时与meanOutOptional的shape相同shape为[A1,...,Ai,1,...,1]Ai后共有j个1与需要norm的轴长度保持相同。FLOAT32、FLOAT16、BFLOAT16ND1-8√workspaceSizeuint64_t*输出返回需要在Device侧申请的workspace大小。-----executoraclOpExecutor**输出返回op执行器包含了算子计算流程。-----返回值aclnnStatus返回状态码具体参见aclnn返回码。第一段接口完成入参校验出现以下场景时报错返回码错误码描述ACLNN_ERR_PARAM_NULLPTR161001传入的input、normalizedShape或out为空指针。ACLNN_ERR_PARAM_INVALID161002input、weightOptional非空时、biasOptional非空时、out、meanOutOptional非空时、rstdOutOptional非空时shape的维度超过8维。normalizedShape的元素个数超过8。input、weightOptional非空时、biasOptional非空时、out、meanOutOptional非空时、rstdOutOptional非空时数据类型不在支持的范围内。normalizedShape维度小于1维。weightOptional非空且shape与normalizedShape不相等。biasOptional非空且shape与normalizedShape不相等。input的维度小于normalizedShape的维度。input的shape与normalizedShape右对齐时对应维度shape不相等。input和out的shape不一致。ACLNN_ERR_INNER_NULLPTR561103normalizedShape的值为[R1,...,Rj]时R1*R2*...*Rj的值大于583705600。aclnnFastLayerNorm参数说明参数名输入/输出描述workspace输入在Device侧申请的workspace内存地址。workspaceSize输入在Device侧申请的workspace大小由第一段接口aclnnFastLayerNormGetWorkspaceSize获取。executor输入op执行器包含了算子计算流程。stream输入指定执行任务的Stream。返回值aclnnStatus返回状态码具体参见aclnn返回码。约束说明input、normalizedShape、weightOptional非空时、biasOptional非空时、out、meanOutOptional非空时或rstdOutOptional非空时的shape不超过8维。确定性计算aclnnFastLayerNorm默认确定性实现。调用示例示例代码如下仅供参考具体编译和执行过程请参考编译与运行样例。#include iostream #include vector #include acl/acl.h #include aclnnop/aclnn_layer_norm.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::vectorint64_t shape) { int64_t shapeSize 1; for (auto i : shape) { shapeSize * i; } return shapeSize; } int Init(int32_t deviceId, aclrtStream* stream) { // 固定写法资源初始化 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 CreateAclTensorMem(const std::vectorT hostData, const std::vectorint64_t shape, void** deviceAddr) { auto size GetShapeSize(shape) * sizeof(T); 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); 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); return 0; } template typename T void aclCreateTensorP(const std::vectorT shape, void** deviceAddr, aclDataType dataType, aclTensor** tensor) { std::vectorint64_t strides(shape.size(), 1); for (int64_t i shape.size() - 2; i 0; i--) { strides[i] shape[i 1] * strides[i 1]; } *tensor aclCreateTensor( shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), *deviceAddr); } template typename T void aclCreateIntArrayP(const std::vectorT hostData, aclIntArray** intArray) { *intArray aclCreateIntArray(hostData.data(), hostData.size()); } int main() { // 1.固定写法device/stream初始化参考acl API手册 int32_t deviceId 0; aclrtStream stream; auto ret Init(deviceId, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(Init acl failed. ERROR: %d\n, ret); return ret); // 2. 构造输入与输出需要根据API的接口自定义构造 std::vectorint64_t xShape {1, 2, 32}; std::vectorint64_t normShape {32}; std::vectorint64_t meanShape {1, 2, 1}; void* xDeviceAddr nullptr; void* weightDeviceAddr nullptr; void* biasDeviceAddr nullptr; void* outDeviceAddr nullptr; void* meanDeviceAddr nullptr; void* rstdDeviceAddr nullptr; aclTensor* x nullptr; aclIntArray* norm nullptr; aclTensor* weight nullptr; aclTensor* bias nullptr; aclTensor* out nullptr; aclTensor* mean nullptr; aclTensor* rstd nullptr; std::vectorfloat xHostData(64, 2.0); std::vectorint64_t normData {32}; std::vectorfloat weightHostData(32, 1.0); std::vectorfloat biasHostData(32, 0.0); std::vectorfloat outHostData(64, 0.0); std::vectorfloat meanHostData(2, 0.0); std::vectorfloat rstdHostData(2, 0.0); double eps 1e-5; ret CreateAclTensorMem(xHostData, xShape, xDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorMem(weightHostData, normShape, weightDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorMem(biasHostData, normShape, biasDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorMem(outHostData, xShape, outDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorMem(meanHostData, meanShape, meanDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorMem(rstdHostData, meanShape, rstdDeviceAddr); CHECK_RET(ret ACL_SUCCESS, return ret); aclCreateTensorP(xShape, xDeviceAddr, aclDataType::ACL_FLOAT, x); aclCreateIntArrayP(normData, norm); aclCreateTensorP(normShape, weightDeviceAddr, aclDataType::ACL_FLOAT, weight); aclCreateTensorP(normShape, biasDeviceAddr, aclDataType::ACL_FLOAT, bias); aclCreateTensorP(xShape, outDeviceAddr, aclDataType::ACL_FLOAT, out); aclCreateTensorP(meanShape, meanDeviceAddr, aclDataType::ACL_FLOAT, mean); aclCreateTensorP(meanShape, rstdDeviceAddr, aclDataType::ACL_FLOAT, rstd); // 3. 调用CANN算子库API需要修改为具体的API名称 uint64_t workspaceSize 0; aclOpExecutor* executor; // 调用aclnnFastLayerNorm第一段接口 ret aclnnFastLayerNormGetWorkspaceSize(x, norm, weight, bias, eps, out, mean, rstd, workspaceSize, executor); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnFastLayerNormGetWorkspaceSize failed. ERROR: %d\n, ret); return ret); 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); } // 调用aclnnFastLayerNorm第二段接口 ret aclnnFastLayerNorm(workspaceAddr, workspaceSize, executor, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnFastLayerNorm failed. ERROR: %d\n, ret); return ret); // 4.固定写法同步等待任务执行结束 ret aclrtSynchronizeStream(stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtSynchronizeStream failed. ERROR: %d\n, ret); return ret); // 5. 获取输出的值将device侧内存上的结果拷贝至host侧需要根据具体API的接口定义修改 auto size GetShapeSize(xShape); std::vectorfloat resultData(size, 0); ret aclrtMemcpy( resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy first result from device to host failed. ERROR: %d\n, ret); return ret); for (int64_t i 0; i size; i) { LOG_PRINT(out result[%ld] is: %f\n, i, resultData[i]); } auto size1 GetShapeSize(meanShape); std::vectorfloat resultData1(size1, 0); ret aclrtMemcpy( resultData1.data(), resultData1.size() * sizeof(resultData1[0]), meanDeviceAddr, size1 * sizeof(resultData1[0]), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy second result from device to host failed. ERROR: %d\n, ret); return ret); for (int64_t i 0; i size1; i) { LOG_PRINT(mean result[%ld] is: %f\n, i, resultData1[i]); } auto size2 GetShapeSize(meanShape); std::vectorfloat resultData2(size2, 0); ret aclrtMemcpy( resultData2.data(), resultData2.size() * sizeof(resultData2[0]), rstdDeviceAddr, size2 * sizeof(resultData2[0]), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy last result from device to host failed. ERROR: %d\n, ret); return ret); for (int64_t i 0; i size2; i) { LOG_PRINT(rstd result[%ld] is: %f\n, i, resultData2[i]); } // 6. 释放aclTensor和aclIntArray需要根据具体API的接口定义修改 aclDestroyTensor(x); aclDestroyIntArray(norm); aclDestroyTensor(weight); aclDestroyTensor(bias); aclDestroyTensor(out); aclDestroyTensor(mean); aclDestroyTensor(rstd); // 7. 释放device资源 aclrtFree(xDeviceAddr); aclrtFree(weightDeviceAddr); aclrtFree(biasDeviceAddr); aclrtFree(outDeviceAddr); aclrtFree(meanDeviceAddr); aclrtFree(rstdDeviceAddr); if (workspaceSize 0) { aclrtFree(workspaceAddr); } aclrtDestroyStream(stream); aclrtResetDevice(deviceId); aclFinalize(); return 0; }【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考