aclnnMhcPostBackward
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：mhc_post基于一系列计算对MHC（Manifold-Constrained Hyper-Connection）架构中上一层输出htouth_{t}^{out}进行Post Mapping，对上一层的输入xjx_j进行ResMapping，然后对二者进行残差连接，得到下一层的输入xl+1x_{l+1}。该算子实现前述过程的反向功能。
-
计算公式：
grad_x=Hlres×grad_outputgrad_h_res=xl×grad_outputTgrad\_x = H_{l}^{res} \times grad\_output\\ grad\_h\_res = x_{l} \times {grad\_output}^{T}
grad_h_out=(grad_output∗(Hlpost.unsqueeze(−1))).sum(dim=−2)grad_h_post=(grad_output∗(hlout.unsqueeze(−2))).sum(dim=−1)grad\_h\_out=({grad\_output} * (H_{l}^{post}.unsqueeze(-1))).sum(dim=-2)\\ grad\_h\_post=({grad\_output} * (h_{l}^{out}.unsqueeze(-2))).sum(dim=-1)
函数原型
算子执行接口为两段式接口，必须先调用“aclnnMhcPostBackwardGetWorkspaceSize”接口获取入参并根据计算流程计算所需workspace大小，再调用“aclnnMhcPostBackward”接口执行计算。
aclnnStatus aclnnMhcPostBackwardGetWorkspaceSize(
const aclTensor *gradOutput,
const aclTensor *x,
const aclTensor *hRes,
const aclTensor *hOut,
const aclTensor *hPost,
aclTensor *gradX,
aclTensor *gradHres,
aclTensor *gradHout,
aclTensor *gradHpost,
uint64_t *workspaceSize,
aclOpExecutor **executor)
aclnnStatus aclnnMhcPostBackward(
void *workspace,
uint64_t workspaceSize,
aclOpExecutor *executor,
const aclrtStream stream)
aclnnMhcPostBackwardGetWorkspaceSize
-
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Tensor gradOutput 输入 待计算的数据，表示网络中MHC层的输入数据 - 不支持空Tensor
FLOAT16、BFLOAT16 ND [B,S,N,D]、[T,N,D] √ x 输入 待计算的数据，表示网络中MHC层的输入数据 - 不支持空Tensor
数据类型与gradOutput一致 ND [B,S,N,D]、[T,N,D] √ hRes 输入 MHC的hRes变换矩阵 - 不支持空Tensor
FLOAT32 ND [B,S,N,N]、[T,N,N] √ hOut 输入 Atten/MLP层的输出 - 不支持空Tensor
数据类型与gradOutput一致 ND [B,S,D]、[T,D] √ hPost 输入 MHC的hPost变换矩阵 - 不支持空Tensor
数据类型与hRes一致 ND [B,S,N]、[T,N] √ gradX 输出 网络中MHC层的输入数据x的梯度 - 数据类型与gradOutput一致 ND [B,S,N,D]、[T,N,D] √ gradHRes 输出 网络中MHC层的输入数据hRes的梯度 - 数据类型与hRes一致 ND [B,S,N,N]、[T,N,N] √ gradHout 输出 网络中MHC层的输入数据hOut的梯度 - 数据类型与hOut一致 ND [B,S,D]、[T,D] √ gradHpost 输出 网络中MHC层的输入数据h_post的梯度 - 数据类型与hPost一致 ND [B,S,N]、[T,N] √ workspaceSize 输出 返回需要在Device侧申请的workspace大小。 - - - - - executor 输出 返回op执行器，包含了算子计算流程。 - - - - - -
返回值：
返回aclnnStatus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：
返回值 错误码 描述 ACLNN_ERR_PARAM_NULLPTR 161001 gradOutput、x、hRes、hOut、hPost存在空指针。 ACLNN_ERR_PARAM_INVALID 161002 gradOutput、x、hRes、hOut、hPost的数据类型不在支持的范围内。 gradOutput、x、hRes、hOut、hPost的shape维度不在支持的范围内。 gradOutput、x、hRes、hOut、hPost的数据类型或shape不匹配。
aclnnMhcPostBackwardGrad
-
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aclnnMhcPostBacwardGetWorkspaceSize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AscendCL stream流。 -
返回值：
返回aclnnStatus状态码，具体参见aclnn返回码。
约束说明
参数说明中维度N的取值目前仅支持4、6和8。
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 <cmath>
#include <cstring>
#include "acl/acl.h"
#include "aclnnop/aclnn_mhc_post_backward.h"
#include "securec.h"
using namespace std;
namespace {
#define CHECK_RET(cond) ((cond) ? true :(false))
#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) {
// Fixed writing method, AscendCL initialization.
auto ret = aclInit(nullptr);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclInit failed. ERROR: %d\n", ret);
return ret;
}
ret = aclrtSetDevice(deviceId);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret);
return ret;
}
ret = aclrtCreateStream(stream);
if (!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 request device side memory.
auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
if (!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);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret);
return ret;
}
// Calculate the strides of continuous tensors.
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 interface to create aclTensor.
*tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
shape.data(), shape.size(), *deviceAddr);
return 0;
}
} // namespace
int main() {
// 1. (Fixed writing method) device/stream initialization. Refer to AscendCL's list of external interfaces.
// Fill in the deviceId based on your actual device.
int32_t deviceId = 0;
aclrtStream stream;
auto ret = Init(deviceId, &stream);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("Init acl failed. ERROR: %d\n", ret);
return ret;
}
LOG_PRINT("ACL environment initialized successfully!\n");
// 2. To construct input and output, it is necessary to customize the construction according to the API interface.
// Example: BSND format (B, S, n, D)
std::vector<int64_t> gradYShape = {1, 1024, 4, 5120}; // BSND
std::vector<int64_t> xShape = {1, 1024, 4, 5120}; // BSND
std::vector<int64_t> hResShape = {1, 1024, 4, 4}; // BSND
std::vector<int64_t> hOutShape = {1, 1024, 5120}; // BSD
std::vector<int64_t> hPostShape = {1, 1024, 4}; // BSn
std::vector<int64_t> gradXShape = {1, 1024, 4, 5120}; // BSND
std::vector<int64_t> gradHresShape = {1, 1024, 4, 4}; // BSND
std::vector<int64_t> gradHoutShape = {1, 1024, 5120}; // BSD
std::vector<int64_t> gradHpostShape = {1, 1024, 4}; // BSn
void *gradYDeviceAddr = nullptr;
void *xDeviceAddr = nullptr;
void *hResDeviceAddr = nullptr;
void *hOutDeviceAddr = nullptr;
void *hPostDeviceAddr = nullptr;
void *gradXDeviceAddr = nullptr;
void *gradHresDeviceAddr = nullptr;
void *gradHoutDeviceAddr = nullptr;
void *gradHpostDeviceAddr = nullptr;
aclTensor *gradYTensor = nullptr;
aclTensor *xTensor = nullptr;
aclTensor *hResTensor = nullptr;
aclTensor *hOutTensor = nullptr;
aclTensor *hPostTensor = nullptr;
aclTensor *gradXTensor = nullptr;
aclTensor *gradHresTensor = nullptr;
aclTensor *gradHoutTensor = nullptr;
aclTensor *gradHpostTensor = nullptr;
int64_t gradYShapeSize = GetShapeSize(gradYShape);
int64_t xShapeSize = GetShapeSize(xShape);
int64_t hResShapeSize = GetShapeSize(hResShape);
int64_t hOutShapeSize = GetShapeSize(hOutShape);
int64_t hPostShapeSize = GetShapeSize(hPostShape);
int64_t gradXShapeSize = GetShapeSize(gradXShape);
int64_t gradHresShapeSize = GetShapeSize(gradHresShape);
int64_t gradHoutShapeSize = GetShapeSize(gradHoutShape);
int64_t gradHpostShapeSize = GetShapeSize(gradHpostShape);
std::vector<aclFloat16> gradYHostData(gradYShapeSize, aclFloatToFloat16(1.0f));
std::vector<aclFloat16> xHostData(xShapeSize, aclFloatToFloat16(1.0f));
std::vector<float> hResHostData(hResShapeSize, 1.0f);
std::vector<aclFloat16> hOutHostData(hOutShapeSize, aclFloatToFloat16(1.0f));
std::vector<float> hPostHostData(hPostShapeSize, 1.0f);
std::vector<aclFloat16> gradXHostData(gradXShapeSize, aclFloatToFloat16(1.0f));
std::vector<float> gradHresHostData(gradHresShapeSize, 1.0f);
std::vector<aclFloat16> gradHoutHostData(gradHoutShapeSize, aclFloatToFloat16(1.0f));
std::vector<float> gradHpostHostData(gradHpostShapeSize, 1.0f);
ret = CreateAclTensor(gradYHostData, gradYShape, &gradYDeviceAddr, aclDataType::ACL_FLOAT16, &gradYTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create x aclTensor.
ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT16, &xTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create hRes aclTensor.
ret = CreateAclTensor(hResHostData, hResShape, &hResDeviceAddr, aclDataType::ACL_FLOAT, &hResTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create hOut aclTensor.
ret = CreateAclTensor(hOutHostData, hOutShape, &hOutDeviceAddr, aclDataType::ACL_FLOAT16, &hOutTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create hPost aclTensor.
ret = CreateAclTensor(hPostHostData, hPostShape, &hPostDeviceAddr, aclDataType::ACL_FLOAT, &hPostTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create gradX aclTensor.
ret = CreateAclTensor(gradXHostData, gradXShape, &gradXDeviceAddr, aclDataType::ACL_FLOAT16, &gradXTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create gradHres aclTensor.
ret = CreateAclTensor(gradHresHostData, gradHresShape, &gradHresDeviceAddr, aclDataType::ACL_FLOAT, &gradHresTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create gradHout aclTensor.
ret = CreateAclTensor(gradHoutHostData, gradHoutShape, &gradHoutDeviceAddr, aclDataType::ACL_FLOAT16, &gradHoutTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// Create gradHpost aclTensor.
ret = CreateAclTensor(gradHpostHostData, gradHpostShape, &gradHpostDeviceAddr, aclDataType::ACL_FLOAT, &gradHpostTensor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
return ret;
}
// 3. Call CANN operator library API.
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
// Call the first interface.
ret = aclnnMhcPostBackwardGetWorkspaceSize(
gradYTensor, xTensor, hResTensor, hOutTensor, hPostTensor, gradXTensor, gradHresTensor, gradHoutTensor,
gradHpostTensor, &workspaceSize, &executor);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclnnMhcPostBackwardGetWorkspaceSize failed. ERROR: %d\n", ret);
return ret;
}
// Apply for device memory based on the workspaceSize calculated from the first interface paragraph.
void *workspaceAddr = nullptr;
if (workspaceSize > 0U) {
ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret);
return ret;
}
}
// Call the second interface.
ret = aclnnMhcPostBackward(workspaceAddr, workspaceSize, executor, stream);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclnnMhcPostBackward failed. ERROR: %d\n", ret);
return ret;
}
// 4. (Fixed writing method) Synchronize and wait for task execution to end.
ret = aclrtSynchronizeStream(stream);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret);
return ret;
}
//gradX, gradHres, gradHout, gradHpost
// 5. Retrieve the output value, copy the gradX from the device side memory to the host side.
auto gradXsize = GetShapeSize(gradXShape);
std::vector<aclFloat16> gradXData(gradXsize, aclFloatToFloat16(0.0f));
ret = aclrtMemcpy(gradXData.data(), gradXData.size() * sizeof(gradXData[0]), gradXDeviceAddr,
gradXsize * sizeof(gradXData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("copy gradX from device to host failed. ERROR: %d\n", ret);
return ret;
}
auto gradHressize = GetShapeSize(gradHresShape);
std::vector<float> gradHresData(gradHressize, 0.0f);
ret = aclrtMemcpy(gradHresData.data(), gradHresData.size() * sizeof(gradHresData[0]), gradHresDeviceAddr,
gradHressize * sizeof(gradHresData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("copy gradHres from device to host failed. ERROR: %d\n", ret);
return ret;
}
auto gradHoutsize = GetShapeSize(gradHoutShape);
std::vector<aclFloat16> gradHoutData(gradHoutsize, aclFloatToFloat16(0.0f));
ret = aclrtMemcpy(gradHoutData.data(), gradHoutData.size() * sizeof(gradHoutData[0]), gradHoutDeviceAddr,
gradHoutsize * sizeof(gradHoutData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("copy gradHout from device to host failed. ERROR: %d\n", ret);
return ret;
}
auto gradHpostsize = GetShapeSize(gradHpostShape);
std::vector<float> gradHpostData(gradHpostsize, 0.0f);
ret = aclrtMemcpy(gradHpostData.data(), gradHpostData.size() * sizeof(gradHpostData[0]), gradHpostDeviceAddr,
gradHpostsize * sizeof(gradHpostData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
if (!CHECK_RET(ret == ACL_SUCCESS)) {
LOG_PRINT("copy gradHpost from device to host failed. ERROR: %d\n", ret);
return ret;
}
// 6. Release resources.
aclDestroyTensor(gradYTensor);
aclDestroyTensor(xTensor);
aclDestroyTensor(hResTensor);
aclDestroyTensor(hOutTensor);
aclDestroyTensor(hPostTensor);
aclDestroyTensor(gradXTensor);
aclDestroyTensor(gradHresTensor);
aclDestroyTensor(gradHoutTensor);
aclDestroyTensor(gradHpostTensor);
aclrtFree(gradYDeviceAddr);
aclrtFree(xDeviceAddr);
aclrtFree(hResDeviceAddr);
aclrtFree(hOutDeviceAddr);
aclrtFree(hPostDeviceAddr);
aclrtFree(gradXDeviceAddr);
aclrtFree(gradHresDeviceAddr);
aclrtFree(gradHoutDeviceAddr);
aclrtFree(gradHpostDeviceAddr);
if (workspaceSize > 0U) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}