aclnnScale
支持的产品型号
- Atlas 推理系列产品。
- Atlas A2训练系列产品/Atlas 800I A2推理产品。
接口原型
每个算子分为两段式接口,必须先调用“aclnnScaleGetWorkspaceSize”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“aclnnScale”接口执行计算。
aclnnstatus aclnnScaleGetWorkspaceSize(const aclTensor *x, const aclTensor *scale, const aclTensor *bias, int64_t axis, int64_t numAxes, bool scaleFromBlob, aclTensor *y, uint64_t *workspaceSize, aclOpExecutor **executor)
aclnnstatus aclnnScale(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
功能描述
计算公式:
若不输入bias,则
若输入bias,则
说明: scale/bias支持跟X的broadcast, scale/bias shape规则 scaleFromBlob is True(axis 转换为正数, numAxes = -1 表示到最后轴) scaleShape = xShape[axis:axis + numAxes] biasShape = xShape[axis:axis + numAxes]
scaleFromBlob is False(axis 转换为正数, numAxes = -1 表示到最前轴) scaleShape must be xShape[axis:axis + rank(scaleShape)] biasShape must be xShape[axis:axis + rank(scaleShape)] ex: scaleFromBlob = True xShape = [a, b, c, d, e, f] axis = 3 numAxes = 2 --> scaleShape = [d, e] xShape = [a, b, c, d, e, f] axis = 3 numAxes = 3 --> scaleShape = [d, e, f] xShape = [a, b, c, d, e, f] axis = 3 numAxes = -1 --> scaleShape = [d, e, f] scaleFromBlob = False xShape = [a, b, c, d, e, f] axis = 3 --> scaleShape = [d, e] xShape = [a, b, c, d, e, f] axis = 3 --> scaleShape = [d, e, f] xShape = [a, b, c, d, e, f] axis = 3 --> scaleShape = [d] xShape = [a, b, c, d, e, f] axis = 3 --> scaleShape = [e] [错误]
aclnnScaleGetWorkspaceSize
参数说明:
- x(aclTensor*, 计算输入): 算子输入的Tensor,Device侧的aclTensor,数据类型支持FLOAT32, FLOAT16、BFLOAT16(Atlas 推理系列产品不支持),支持非连续的Tensor,数据格式支持ND.
- scale(aclTensor*, 计算输入): 算子输入的Tensor,Device侧的aclTensor,数据类型支持FLOAT32, FLOAT16、BFLOAT16(Atlas 推理系列产品不支持),数据类型需要与x的数据类型相同,支持非连续的Tensor,数据格式支持ND.shape满足broadcast要求,见功能描述章节说明.
- bias(aclTensor*, 计算输入): 算子输入的Tensor,Device侧的aclTensor,数据类型支持FLOAT32, FLOAT16、BFLOAT16(Atlas 推理系列产品不支持),不为空时数据类型需要与scale的数据类型相同,支持非连续的Tensor,数据格式支持ND.shape与scale保持一致.
- axis(int64_t, 计算输入) host侧INT64类型,指定进行scale的起始轴, 取值范围 [-x_rank, x_rank)(x_rank表示x的shape维度).
- numAxes(int64_t, 计算输入) host侧INT64类型,指定进行scale的轴长度, 取值范围 >= -1, numAxes = -1, 表示从axis轴开始scale到最后一轴.
- scaleFromBlob(bool, 计算输入) host侧BOOL类型,指定要scaleFromBlob类型, True: scale from blob, 使用numAxes + axis进行scale, False: scale from input scale, 从axis开始 scale input scale长度, 忽略numAxes取值.
- y(aclTensor*, 计算输出): 输出Tensor,shape维度和x保持一致,Device侧的aclTensor,数据类型支持FLOAT32, FLOAT16、BFLOAT16(Atlas 推理系列产品不支持),数据类型需要与x的数据类型相同,支持非连续的Tensor,数据格式支持ND.
- workspaceSize(uint64_t*, 出参): 返回需要在Device侧申请的workspace大小.
- executor(aclOpExecutor**, 出参): 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.
返回161001(ACLNN_ERR_PARAM_NULLPTR):1. 传入的x、scale、y是空指针。 返回161002(ACLNN_ERR_PARAM_INVALID):1. x的数据类型不在支持的范围之内。 2. bias不为空时,bias与scale的数据类型不一致。 3. scale与x的数据类型不一致。 4. y与x的数据类型不一致。 5. x和y的shape不一致。 6. bias不为空时,bias与scale的shape不一致。 7. x和scale的shape维度大于8. 8. axis的取值不在[-x_rank, x_rank)范围内。 9. numAxes的取值小于-1。 10. scaleFromBlob为True,numAxes等于0且scale的shape不为[1]。 11. axis转换为正数之后与numAxes相加,大于x_rank。 12. scale的shape与预期不符(预期shape推导参考功能描述)。
aclnnScale
参数说明:
- workspace(void *, 入参):在Device侧申请的workspace内存地址。
- workspaceSize(uint64_t, 入参):在Device侧申请的workspace大小,由第一段接口aclnnScaleGetWorkspaceSize获取。
- executor(aclOpExecutor *, 入参):op执行器,包含了算子计算流程。
- stream(aclrtStream, 入参):指定执行任务的 AscendCL Stream流。
返回值:
aclnnStatus:返回状态码,具体参见aclnn返回码。
约束与限制
Atlas 推理系列产品不支持scale、offset及输入为bf16。
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/level2/aclnn_scale.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) {
// 固定写法,AscendCL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 CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
aclDataType dataType, aclTensor** tensor) {
auto size = GetShapeSize(shape) * sizeof(T);
// 调用aclrtMalloc申请device侧内存
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);
// 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
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);
// 计算连续tensor的strides
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];
}
// 调用aclCreateTensor接口创建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. (固定写法)device/stream初始化,参考AscendCL对外接口列表
// 根据自己的实际device填写deviceId
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::vector<int64_t> selfShape = {4, 2};
std::vector<int64_t> tensor1Shape = {4};
std::vector<int64_t> tensor2Shape = {4};
std::vector<int64_t> outShape = {4, 2};
void* selfDeviceAddr = nullptr;
void* tensor1DeviceAddr = nullptr;
void* tensor2DeviceAddr = nullptr;
void* outDeviceAddr = nullptr;
aclTensor* self = nullptr;
aclTensor* tensor1 = nullptr;
aclTensor* tensor2 = nullptr;
aclTensor* out = nullptr;
std::vector<float> selfHostData = {0, 1, 2, 3, 4, 5, 6, 7};
std::vector<float> tensor1HostData = {2, 2, 2, 2, 2, 2, 2, 2};
std::vector<float> tensor2HostData = {2, 2, 2, 2, 2, 2, 2, 2};
std::vector<float> outHostData = {0, 0, 0, 0, 0, 0, 0, 0};
int64_t axis = 0;
int64_t numAxes = 1;
bool fromBlom = true;
// 创建self aclTensor
ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建tensor1 aclTensor
ret = CreateAclTensor(tensor1HostData, tensor1Shape, &tensor1DeviceAddr, aclDataType::ACL_FLOAT, &tensor1);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建tensor2 aclTensor
ret = CreateAclTensor(tensor2HostData, tensor2Shape, &tensor2DeviceAddr, aclDataType::ACL_FLOAT, &tensor2);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建out aclTensor
ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 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aclnnScale第一段接口
ret = aclnnScaleGetWorkspaceSize(self, tensor1, tensor2, axis, numAxes, fromBlom, out, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnScaleGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
// 根据第一段接口计算出的workspaceSize申请device内存
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aclnnScale第二段接口
ret = aclnnScale(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnScale 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(outShape);
std::vector<float> 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 resultData from device to host failed. ERROR: %d\n", ret);
return ret);
for (int64_t i = 0; i < size; i++) {
LOG_PRINT("resultData[%ld] is: %f\n", i, resultData[i]);
}
// 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
aclDestroyTensor(self);
aclDestroyTensor(tensor1);
aclDestroyTensor(tensor2);
aclDestroyTensor(out);
// 7.释放device资源,需要根据具体API的接口定义修改
aclrtFree(selfDeviceAddr);
aclrtFree(tensor1DeviceAddr);
aclrtFree(tensor2DeviceAddr);
aclrtFree(outDeviceAddr);
if (workspaceSize > 0) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}