aclnnSwinTransformerLnQkvQuant
支持的产品型号
- Atlas 推理系列产品(Ascend 310P处理器)。
接口原型
每个算子分为两段式接口,必须先调用“aclnnSwinTransformerLnQkvQuantGetWorkspaceSize”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“aclnnSwinTransformerLnQkvQuant”接口执行计算。
aclnnStatus aclnnSwinTransformerLnQkvQuantGetWorkspaceSize(const aclTensor *x, const aclTensor *gamma, aclTensor *beta, aclTensor *weight, aclTensor *bias, aclTensor *quantScale, aclTensor *quantOffset, aclTensor *dequantScale, int64_t headNum, int64_t seqLength, float epsilon, int64_t oriHeight, int64_t oriWeight, int64_t hWinSize, int64_t wWinSize, bool weightTranspose, aclTensor *queryOutput, aclTensor *keyOutput, aclTensor *valueOutput, uint64_t *workspaceSize, aclOpExecutor **executor)
aclnnStatus aclnnSwinTransformerLnQkvQuant(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
功能描述
算子功能:Swin Transformer 网络模型 完成 Q K V 的计算。
计算公式:
q/k/v = (Quant(Layernorm(x).transpose) * weight).dequant.transpose.split 其中,weight 是 Q K V 三个矩阵权重的拼接。
aclnnSwinTransformerLnQkvQuantGetWorkspaceSize
- 参数说明:
- x(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持FLOAT16。数据格式支持ND。只支持维度为[B,S,H],其中B只支持[1,32],S需等于oriHeight * oriWeight,H = headNum * seqLength且小于等于1024。
- gamma(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持FLOAT16。数据格式支持ND。维度需为[H]。
- beta(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持FLOAT16。数据格式支持ND。维度需为[H]。
- weight(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持INT8。数据格式支持ND。维度需为[3 * H, H]。
- bias(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持INT32。数据格式支持ND。维度需为[3 * H]。
- quantScale(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持FLOAT16。数据格式支持ND。维度需为[H]。
- quantOffset(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持FLOAT16。数据格式支持ND。维度需为[H]。
- dequantScale(const aclTensor*,计算输入): Device侧的aclTensor,数据类型支持UINT64。数据格式支持ND。维度需为[3 * H]。
- headNum(int,计算输入): 通道数;只支持[1,32]范围;
- seqLength(int,计算输入): 通道深度。只支持32/64两种;
- epsilon(float,计算输入): layernrom 计算除0保护值;为了保证精度,建议小于等于1e-4;
- oriHeight(int,计算输入): layernrom 中S轴transpose的维度;oriHeight*oriWeight需等于输入x的第二维S的大小,且为hWinSize的整数倍;
- oriWeight(int,计算输入): layernrom 中S轴transpose的维度;oriHeight*oriWeight需等于输入x的第二维S的大小,且为wWinSize的整数倍;
- hWinSize(int,计算输入): 窗大小;支持范围[7-32];
- wWinSize(int,计算输入): 窗大小;支持范围[7-32];
- weightTranspose(bool,计算输入): weight矩阵需要转置,当前不支持不转置场景;
- queryOutput(const aclTensor*, 计算输出):Device侧的aclTensor,数据类型支持FLOAT16数据格式支持ND。
- keyOutput(const aclTensor*, 计算输出):Device侧的aclTensor,数据类型支持FLOAT16数据格式支持ND。
- valueOutput(const aclTensor*, 计算输出):Device侧的aclTensor,数据类型支持FLOAT16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。
161001(ACLNN_ERR_PARAM_NULLPTR):1. 传入的输入tensor是空指针。 161002(ACLNN_ERR_PARAM_INVALID):1. 输入或输出参数的数据类型/数据格式不在支持的范围内。
aclnnSwinTransformerLnQkvQuant
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aclnnSwinTransformerLnQkvQuantGetWorkspaceSize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
- seqLength 只支持32/64
- oriHeight * oriWeight = 输入x Tensor的第二维度,且oriHeight为hWinSize的整数倍,oriWeight为wWinSize的整数倍
- hWinSize和wWinSize 只支持7 - 32
- 输入x Tensor的第一维度B只支持1 - 32
- weight需要转置
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_swin_transformer_ln_qkv_quant.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 = {1, 49, 32};
std::vector<int64_t> gammaShape = {32};
std::vector<int64_t> weightShape = {32*3, 32};
std::vector<int64_t> biasShape = {3 * 32};
std::vector<int64_t> outShape = {1,1,49, 32};
void* xDeviceAddr = nullptr;
void* gammaDeviceAddr = nullptr;
void* betaDeviceAddr = nullptr;
void* weightDeviceAddr = nullptr;
void* biasDeviceAddr = nullptr;
void* scaleDeviceAddr = nullptr;
void* offsetDeviceAddr = nullptr;
void* dequantDeviceAddr = nullptr;
void* outqDeviceAddr = nullptr;
void* outkDeviceAddr = nullptr;
void* outvDeviceAddr = nullptr;
aclTensor* x = nullptr;
aclTensor* gamma = nullptr;
aclTensor* beta = nullptr;
aclTensor* weight = nullptr;
aclTensor* bias = nullptr;
aclTensor* quantScale = nullptr;
aclTensor* quantOffset = nullptr;
aclTensor* dequantScale = nullptr;
aclTensor* queryOutput = nullptr;
aclTensor* keyOutput = nullptr;
aclTensor* valueOutput = nullptr;
std::vector<uint16_t> selfHostData(49*32, 0x1);
std::vector<int32_t> biasHostData(3*32, 0x1);
std::vector<uint16_t> gammaHostData(32, 0x1);
std::vector<uint16_t> betaHostData(32, 0x1);
std::vector<int8_t> weightHostData(3*32*32, 0x1);
std::vector<uint16_t> scaleHostData(32, 0x1);
std::vector<uint16_t> offsetHostData(32, 0x1);
std::vector<uint64_t> dequantHostData(3*32, 0x1);
std::vector<uint16_t> outqHostData(49*32, 0x0);
std::vector<uint16_t> outkHostData(49*32, 0x0);
std::vector<uint16_t> outvHostData(49*32, 0x0);
// 创建self aclTensor
ret = CreateAclTensor(selfHostData, selfShape, &xDeviceAddr, aclDataType::ACL_FLOAT16, &x);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(gammaHostData, gammaShape, &gammaDeviceAddr, aclDataType::ACL_FLOAT16, &gamma);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(betaHostData, gammaShape, &betaDeviceAddr, aclDataType::ACL_FLOAT16, &beta);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(weightHostData, weightShape, &weightDeviceAddr, aclDataType::ACL_INT8, &weight);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(biasHostData, biasShape, &biasDeviceAddr, aclDataType::ACL_INT32, &bias);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(scaleHostData, gammaShape, &scaleDeviceAddr, aclDataType::ACL_FLOAT16, &quantScale);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(offsetHostData, gammaShape, &offsetDeviceAddr, aclDataType::ACL_FLOAT16, &quantOffset);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(dequantHostData, biasShape, &dequantDeviceAddr, aclDataType::ACL_UINT64, &dequantScale);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建out aclTensor
ret = CreateAclTensor(outqHostData, outShape, &outqDeviceAddr, aclDataType::ACL_FLOAT16, &queryOutput);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(outkHostData, outShape, &outkDeviceAddr, aclDataType::ACL_FLOAT16, &keyOutput);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(outvHostData, outShape, &outvDeviceAddr, aclDataType::ACL_FLOAT16, &valueOutput);
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;
float epsilon = 0.0001;
int64_t oriHeight = 7;
int64_t oriWeight = 7;
int64_t hWinSize = 7;
int64_t wWinSize = 7;
int64_t headNum = 1;
int64_t seqLength = 32;
bool weightTranspose = true;
// 调用aclnnSwinTransformerLnQkvQuant第一段接口
ret = aclnnSwinTransformerLnQkvQuantGetWorkspaceSize(x,gamma,beta,weight, bias, quantScale, quantOffset, dequantScale, headNum, seqLength, epsilon, oriHeight, oriWeight, hWinSize, wWinSize, weightTranspose, queryOutput, keyOutput, valueOutput, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSwinTransformerLnQkvQuantGetWorkspaceSize 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aclnnSwinTransformerLnQkvQuant第二段接口
ret = aclnnSwinTransformerLnQkvQuant(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSwinTransformerLnQkvQuant 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<uint16_t> resultData(size, 0);
ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outqDeviceAddr,size * sizeof(resultData[0]), 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);
// 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
aclDestroyTensor(x);
aclDestroyTensor(queryOutput);
// 7. 释放device资源,需要根据具体API的接口定义修改
aclrtFree(xDeviceAddr);
aclrtFree(outqDeviceAddr);
if (workspaceSize > 0) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}