aclnnWeightQuantBatchMatmul
该接口后续版本会废弃,请使用aclnnWeightQuantBatchMatmulV2接口。
支持的产品型号
- Atlas A2训练系列产品/Atlas 800I A2推理产品。
接口原型
- 两段式接口,必须先调用“aclnnWeightQuantBatchMatmulGetWorkspaceSize”接口获取入参并根据计算流程计算所需workspace大小,再调用“aclnnWeightQuantBatchMatmul”接口执行计算。
aclnnStatus aclnnWeightQuantBatchMatmulGetWorkspaceSize(const aclTensor *x1, const aclTensor *x2, const aclTensor *diagonalMatrix, const aclTensor *deqOffset, const aclTensor *deqScale, const aclTensor *addOffset, const aclTensor *mulScale, const aclTensor *bias, bool transposeX1, bool transposeX2, float antiquantScale, float antiquantOffset, aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
aclnnStatus aclnnWeightQuantBatchMatmul(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
功能描述
- 算子功能:伪量化用于对self * mat2(matmul/batchmatmul)中的mat2进行量化。
- 计算公式:
aclnnWeightQuantBatchMatmulGetWorkspaceSize
参数说明
- x1(aclTensor*, 计算输入):公式中的输入
self
,数据类型支持FLOAT16,数据格式支持ND。不支持非连续的Tensor。维度仅支持二维不支持batch轴,与x2需满足broadcast关系。 - x2(aclTensor*, 计算输入):经处理能得到公式中的输入
mat2
,数据类型支持INT8,数据格式支持ND。不支持非连续的Tensor。维度仅支持二维不支持batch轴,但与x1需满足broadcast关系。 - diagonalMatrix(aclTensor*, 计算输入):对x2反量化得到公式中的输入
mat2
,数据类型支持INT8,数据格式支持ND。不支持非连续的Tensor。维度固定为二维,shape为(32, 32),为单位矩阵,m > 64时不参与计算且可以为空。 - deqOffset(aclTensor*, 计算输入):对x2反量化得到公式中的输入
mat2
,由addOffset、antiquantOffset、antiquantScale计算得到,计算方式见示例代码,数据类型支持INT32,数据格式支持ND。不支持非连续的Tensor。shape支持 1 或者 n 或者(1, 1)或者(1, n)或者(n, 1),需和x2满足broadcast关系。m > 64时不参与计算且可以为空。 - deqScale(aclTensor*, 计算输入):对x2反量化得到公式中的输入
mat2
,由接口aclnnTransQuantParam计算得到,计算方式见示例代码,数据类型支持UINT64,数据格式支持ND。不支持非连续的Tensor。shape支持 1 或者 n 或者(1, 1) 或者(1, n) 或者(n, 1),需和x2满足broadcast关系。m > 64时不参与计算且可以为空。 - addOffset(aclTensor*, 计算输入):对x2反量化得到公式中的输入
mat2
,数据类型支持FLOAT16,数据格式支持ND。不支持非连续的Tensor。shape支持 1 或者 n 或者(1, 1)或者(1, n)或者(n, 1),需和x2满足broadcast关系。m < 64时不参与计算, 任意情况都可以为空。 - mulScale(aclTensor*, 计算输入):对x2反量化得到公式中的输入
mat2
,数据类型支持FLOAT16,数据格式支持ND。不支持非连续的Tensor。shape支持 1 或者 n 或者(1, 1)或者(1, n)或者(n, 1),需和x2满足broadcast关系。m < 64时不参与计算, 任意情况都可以为空。 - bias(aclTensor*, 计算输入):公式中的输入
bias
,数据类型支持FLOAT,数据格式支持ND。不支持非连续的Tensor。维度为一维且值等于N,可以为空。 - transposeX1(bool, 计算输入):用于描述x1是否转置。
- transposeX2(bool, 计算输入):用于描述x2是否转置。
- antiquantScale(float, 计算输入):对x2反量化得到公式中的输入
mat2
。 - antiquantOffset(float, 计算输入):对x2反量化得到公式中的输入
mat2
。 - out(aclTensor*, 计算输出):公式中的
result
,数据类型支持FLOAT16和INT8,且数据类型需要是x1与x2推导之后可转换的数据类型,shape需要是x1与x2 broadcast之后的shape。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。
- x1(aclTensor*, 计算输入):公式中的输入
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161001 (ACLNN_ERR_PARAM_NULLPTR): 1. 传入的x1或x2,diagonalMatrix(m < 64时),deqOffset(m < 64时),deqScale(m < 64时)是空指针。 返回161002 (ACLNN_ERR_PARAM_INVALID): 1. 传入非空tensor的数据类型不在支持的范围之内。
aclnnWeightQuantBatchMatmul
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aclnnWeightQuantBatchMatmulGetWorkspaceSize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
无
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 "acl/acl.h"
#include "aclnnop/aclnn_trans_quant_param.h"
#include "aclnnop/aclnn_cast.h"
#include "aclnnop/aclnn_weight_quant_batch_matmul.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> x1Shape = {128, 4};
std::vector<int64_t> x2Shape = {4, 4};
std::vector<int64_t> addOffsetShape = {4};
std::vector<int64_t> mulScaleShape = {4};
std::vector<int64_t> diagonalMatrixShape = {32, 32};
std::vector<int64_t> deqOffsetShape = {4};
std::vector<int64_t> deqScaleShape = {4};
std::vector<int64_t> outShape = {128, 4};
void* x1DeviceAddr = nullptr;
void* x2DeviceAddr = nullptr;
void* addOffsetDeviceAddr = nullptr;
void* mulScaleDeviceAddr = nullptr;
void* diagonalMatrixDeviceAddr = nullptr;
void* deqOffsetDeviceAddr = nullptr;
void* deqScaleDeviceAddr = nullptr;
void* outDeviceAddr = nullptr;
void* x1Fp16DeviceAddr = nullptr;
void* addOffsetFp16DeviceAddr = nullptr;
void* mulScaleFp16DeviceAddr = nullptr;
void* outFp16DeviceAddr = nullptr;
std::vector<float> x1HostData = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
std::vector<float> x2HostData = {1, 1, 1, 1,
1, 1, 1, 1,
1, 1, 1, 1,
1, 1, 1, 1};
std::vector<float> outHostData = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
bool transposeX1 = false;
bool transposeX2 = false;
float antiquantOffset = 0;
float antiquantScale = 1;
std::vector<float> addOffsetHostData = {1.0, 1.0, 1.0, 0.0};
float* addOffsetDate = addOffsetHostData.data();
uint64_t addOffsetSize = 4;
std::vector<float> mulScaleHostData = {2.0, 2.0, 1.0, 1.0};
float* mulScaleDate = mulScaleHostData.data();
uint64_t mulScaleSize = 4;
// diagonalMatrixData
uint64_t n = 32;
uint64_t diagonalMatrixSize = n*n;
int8_t *diagonalMatrixData = (int8_t *)calloc(diagonalMatrixSize, sizeof(int32_t));
for (int64_t i = 0; i < n; i++) {
diagonalMatrixData[i * n + i] = 1;
}
std::vector<int8_t> diagonalMatrixHostData(diagonalMatrixData, diagonalMatrixData + diagonalMatrixSize);
// Get deqOffset
uint64_t deqOffsetSize = addOffsetSize;
int32_t *deqOffsetData = (int32_t *)calloc(deqOffsetSize, sizeof(int32_t));
for (int64_t i = 0; i < deqOffsetSize; i++) {
deqOffsetData[i] = static_cast<int32_t>(round(addOffsetDate[i] / antiquantScale - antiquantOffset));
}
std::vector<int32_t> deqOffsetHostData(deqOffsetData, deqOffsetData + deqOffsetSize);
// Get deqScale
uint64_t deqScaleSize = mulScaleSize;
uint64_t *deqScaleData = (uint64_t *)calloc(deqScaleSize, sizeof(uint64_t));
for (int64_t i = 0; i < deqScaleSize; i++) {
mulScaleDate[i] = mulScaleDate[i] * antiquantScale;
}
std::vector<uint64_t> deqScaleHostData(deqScaleData, deqScaleData + deqScaleSize);
// creat aclTensor
aclTensor* x1 = nullptr;
aclTensor* x2 = nullptr;
aclTensor* addOffset = nullptr;
aclTensor* mulScale = nullptr;
aclTensor* diagonalMatrix = nullptr;
aclTensor* deqOffset = nullptr;
aclTensor* deqScale = nullptr;
aclTensor* out = nullptr;
aclTensor* x1Fp16 = nullptr;
aclTensor* addOffsetFp16 = nullptr;
aclTensor* mulScaleFp16 = nullptr;
aclTensor* outFp16 = nullptr;
// 创建x1 aclTensor
ret = CreateAclTensor(x1HostData, x1Shape, &x1DeviceAddr, aclDataType::ACL_FLOAT, &x1);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建x1Fp16 aclTensor
ret = CreateAclTensor(x1HostData, x1Shape, &x1Fp16DeviceAddr, aclDataType::ACL_FLOAT16, &x1Fp16);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建x2 aclTensor
ret = CreateAclTensor(x2HostData, x2Shape, &x2DeviceAddr, aclDataType::ACL_INT8, &x2);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建addOffset aclTensor
ret = CreateAclTensor(addOffsetHostData, addOffsetShape, &addOffsetDeviceAddr, aclDataType::ACL_FLOAT, &addOffset);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建addOffsetFp16 aclTensor
ret = CreateAclTensor(addOffsetHostData, addOffsetShape, &addOffsetFp16DeviceAddr, aclDataType::ACL_FLOAT16, &addOffsetFp16);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建mulScale aclTensor
ret = CreateAclTensor(mulScaleHostData, mulScaleShape, &mulScaleDeviceAddr, aclDataType::ACL_FLOAT, &mulScale);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建mulScaleFp16 aclTensor
ret = CreateAclTensor(mulScaleHostData, mulScaleShape, &mulScaleFp16DeviceAddr, aclDataType::ACL_FLOAT16, &mulScaleFp16);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建diagonalMatrix aclTensor
ret = CreateAclTensor(diagonalMatrixHostData, diagonalMatrixShape, &diagonalMatrixDeviceAddr, aclDataType::ACL_INT8, &diagonalMatrix);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建deqOffset aclTensor
ret = CreateAclTensor(deqOffsetHostData, deqOffsetShape, &deqOffsetDeviceAddr, aclDataType::ACL_INT32, &deqOffset);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 创建deqScale aclTensor
ret = CreateAclTensor(deqScaleHostData, deqScaleShape, &deqScaleDeviceAddr, aclDataType::ACL_UINT64, &deqScale);
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);
// 创建outFp16 aclTensor
ret = CreateAclTensor(outHostData, outShape, &outFp16DeviceAddr, aclDataType::ACL_FLOAT16, &outFp16);
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
// aclnnWeightQuantBatchMatmul接口调用示例
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
// aclnn cast fp16
//x1
ret = aclnnCastGetWorkspaceSize(x1, aclDataType::ACL_FLOAT16, x1Fp16, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCastGetWorkspaceSize 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);
}
ret = aclnnCast(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCast failed. ERROR: %d\n", ret); return ret);
ret = aclrtSynchronizeStream(stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
// addOffset
ret = aclnnCastGetWorkspaceSize(addOffset, aclDataType::ACL_FLOAT16, addOffsetFp16, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCastGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
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);
}
ret = aclnnCast(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCast failed. ERROR: %d\n", ret); return ret);
ret = aclrtSynchronizeStream(stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
// mulScale
ret = aclnnCastGetWorkspaceSize(mulScale, aclDataType::ACL_FLOAT16, mulScaleFp16, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCastGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
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);
}
ret = aclnnCast(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCast failed. ERROR: %d\n", ret); return ret);
ret = aclrtSynchronizeStream(stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
// 调用aclnnWeightQuantBatchMatmul第一段接口
ret = aclnnWeightQuantBatchMatmulGetWorkspaceSize(x1Fp16, x2, diagonalMatrix, deqOffset, deqScale, addOffsetFp16, mulScaleFp16, nullptr, transposeX1, transposeX2, antiquantScale, antiquantOffset, outFp16, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnWeightQuantBatchMatmulGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
// 根据第一段接口计算出的workspaceSize申请device内存
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aclnnWeightQuantBatchMatmul第二段接口
ret = aclnnWeightQuantBatchMatmul(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnWeightQuantBatchMatmul 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);
// fp16 to fp 32
ret = aclnnCastGetWorkspaceSize(outFp16, aclDataType::ACL_FLOAT, out, &workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCastGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
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);
}
ret = aclnnCast(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCast failed. ERROR: %d\n", ret); return ret);
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 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. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
aclDestroyTensor(x1);
aclDestroyTensor(x1Fp16);
aclDestroyTensor(x2);
aclDestroyTensor(addOffset);
aclDestroyTensor(addOffsetFp16);
aclDestroyTensor(mulScaleFp16);
aclDestroyTensor(mulScale);
aclDestroyTensor(out);
aclDestroyTensor(outFp16);
// 7. 释放device资源,需要根据具体API的接口定义修改
aclrtFree(x1DeviceAddr);
aclrtFree(x2DeviceAddr);
aclrtFree(addOffsetDeviceAddr);
aclrtFree(deqScaleDeviceAddr);
aclrtFree(mulScaleDeviceAddr);
aclrtFree(diagonalMatrixDeviceAddr);
aclrtFree(deqOffsetDeviceAddr);
aclrtFree(addOffsetDeviceAddr);
aclrtFree(outDeviceAddr);
aclrtFree(x1Fp16DeviceAddr);
aclrtFree(addOffsetFp16DeviceAddr);
aclrtFree(mulScaleFp16DeviceAddr);
aclrtFree(outFp16DeviceAddr);
if (workspaceSize > 0) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
free(diagonalMatrixData);
free(deqOffsetData);
free(deqScaleData);
return 0;
}