aclnnQuantMatmulAllReduce
支持的产品型号
- Atlas A2训练系列产品/Atlas 800I A2推理产品。
说明： 使用该接口时，请确保驱动固件包和CANN包都为配套的8.0.RC2版本或者配套的更高版本，否则将会引发报错，比如BUS ERROR等。
接口原型
每个算子分为两段式接口，必须先调用“aclnnQuantMatmulAllReduceGetWorkspaceSize”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“aclnnQuantMatmulAllReduce”接口执行计算。
aclnnStatus aclnnQuantMatmulAllReduceGetWorkspaceSize(const aclTensor *x1, const aclTensor *x2, const aclTensor *bias, const aclTensor *x3, const aclTensor *dequantScale, const char* group, const char *reduceOp, int64_t commTurn, int64_t streamMode, const aclTensor *output, uint64_t *workspaceSize, aclOpExecutor **executor);aclnnStatus aclnnQuantMatmulAllReduce(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream);
功能描述
- 算子功能：对量化后的入参x1、x2进行matmul计算后，接着进行dequant计算，接着与x3进行add操作，最后做all_reduce计算。
- 计算公式：
aclnnQuantMatmulAllReduceGetWorkspaceSize
参数说明：
- x1（aclTensor*, 计算输入)：公式中的输入x1，数据类型支持INT8，数据格式支持ND，不支持非连续输入。Device侧的aclTensor，mm左矩阵，当前版本仅支持二维或者三维输入。
- x2（aclTensor*, 计算输入)：公式中的输入x2，数据类型支持INT8，Device侧的aclTensor，mm右矩阵。当x2的Format为FRACTAL_NZ时，配合aclnnCalculateMatmulWeightSizeV2和aclnnTransMatmulWeight完成输入ND到NZ的转换，非连续的tensor仅支持transpose场景。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：数据格式支持ND（当前版本仅支持二维输入）和FRACTAL_NZ格式（当前版本仅支持四维输入）。
- bias（aclTensor*, 计算输入)：公式中的输入bias，数据类型支持INT32，数据格式支持ND。Device侧的aclTensor，可选，可为空。当前版本仅支持一维输入。
- x3（aclTensor*, 计算输入)：公式中的输入x3，数据格式支持ND。Device侧的aclTensor，matmul计算后的add计算，维度与output一致，可选，可为空。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：数据类型支持FLOAT16、BFLOAT16。
- dequantScale（aclTensor*, 计算输入)：公式中的输入dequantScale，数据格式支持ND。matmul计算后的去量化系数，可选，可为空，shape在per-tensor场景为[1]，per-channel场景为[n]/[1, n]。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：数据类型支持INT64、UINT64、BFLOAT16。
- group(char*, 计算输入)：通信域名称，数据类型支持String，通过Hccl提供的接口获取：extern HcclResult HcclGetCommName(HcclComm comm, char* commName); commName即为group。
- reduceOp(char*, 计算输入)：reduce操作类型，数据类型支持String，当前版本仅支持"sum"。
- commTurn(int64_t, 计算输入)：通信数据切分数，数据类型支持int64_t，即总数据量/单次通信量。当前版本仅支持输入为0。
- streamMode(int64_t, 计算输入)：Host侧的整型，数据类型支持int64_t，AscendCL流模式的枚举，当前只支持枚举值1。
- output(aclTensor *, 输出)：公式中的输出output。Device侧的aclTensor，计算+通信的结果。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：数据类型支持FLOAT16、BFLOAT16。
- 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：
返回aclnnStatus状态码，具体参见aclnn返回码。
aclnnQuantMatmulAllReduce
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aclnnQuantMatmulAllReduceGetWorkspaceSize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：
返回aclnnStatus状态码，具体参见aclnn返回码。
约束与限制
- 增量场景不使能MC2，全量场景使能MC2。
- 输入x1可为二维或者三维，其维度为(b, s, k)或者(m, k)。
- x2必须是二维。其维度为(k, n)，k轴满足mm算子入参要求，k轴相等。
- 不支持x1、x2为空矩阵。
- m大小不超过2147483647，x1与x2的最后一维大小不超过65535，x1的最后一维指k，x2的最后一维指转置时的k或非转置时的n。bias若非空，维度为(n)。x3若非空，shape大小与output相等。
- 当输入x1维度为(b, s, k)时，输出output维度为(b, s, n)，当输入x1维度为(m, k)时，输出output维度为(m, n)。
- 传入的x1、x2、dequantScale或者output不为空指针。
- x1和x2、dequantScale、output、bias（非空场景）、x3（非空场景）的数据类型和数据格式需要在支持的范围之内。
- 若输出output类型为FLOAT16，dequantScale的类型为INT64、UINT64；若输出output类型为BFLOAT16，dequantScale的类型为BFLOAT16，x3的类型为BFLOAT16。
- streamMode的数据在可选范围内，目前仅支持1。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：只支持x2矩阵转置/不转置，x1矩阵不支持转置场景。
- Atlas A2训练系列产品/Atlas 800I A2推理产品：支持1、2、4、8卡，并且仅支持hccs链路all mesh组网。
- 一个模型中的通算融合MC2算子，仅支持相同通信域。
调用示例
Atlas A2训练系列产品/Atlas 800I A2推理产品：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 <thread>
#include "aclnnop/aclnn_trans_matmul_weight.h"
#include "aclnnop/aclnn_quant_matmul_all_reduce.h"
int ndev = 8;
#define ACL_CHECK(ret) \
do { \
auto retcode = ret; \
if (retcode != ACL_SUCCESS) { \
printf("[ERROR] acl interface return err %s:%d, retcode: %d \n", __FILE__, __LINE__, retcode); \
return retcode; \
} \
} while (0)
#define CHECK_RET(cond, return_expr) \
do { \
if (!(cond)) { \
return_expr; \
} \
} while (0)
#define LOG_PRINT(message, ...) \
do { \
printf(message, ##__VA_ARGS__); \
} while (0)
struct Args {
uint32_t rankId;
HcclComm hcclComm;
aclrtStream stream;
aclrtContext context;
std::string format;
};
int64_t GetShapeSize(const std::vector<int64_t> &shape) {
int64_t shapeSize = 1;
for (auto i: shape) {
shapeSize *= i;
}
return shapeSize;
}
template<typename T>
int CreateWeightNzAclTensor(const std::vector<T> &hostData, const std::vector<int64_t> &shape, void **deviceAddr,
aclDataType dataType, aclTensor **tensor, Args &args) {
auto size = GetShapeSize(shape) * sizeof(T);
const aclIntArray *mat2Size = aclCreateIntArray(shape.data(), shape.size());
auto ret = aclnnCalculateMatmulWeightSizeV2(mat2Size, ACL_INT8, &size);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCalculateMatmulWeightSizeV2 failed. ERROR: %d\n", ret); return ret);
auto tensorSize = size * sizeof(T);
// 调用aclrtMalloc申请device内存
ret = aclrtMalloc(deviceAddr, tensorSize, 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);
uint64_t transWorkspaceSize;
aclOpExecutor *executor;
void *transWorkspaceAddr = nullptr;
ret = aclnnTransMatmulWeightGetWorkspaceSize(*tensor, &transWorkspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS && transWorkspaceSize > 0,
printf("[ERROR] aclnnTransMatmulWeightGetWorkspaceSize failed. ret = %d \n", ret); return ret);
ACL_CHECK(aclrtMalloc(&transWorkspaceAddr, transWorkspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
ret = aclnnTransMatmulWeight(transWorkspaceAddr, transWorkspaceSize, executor, args.stream);
CHECK_RET(ret == ACL_SUCCESS, printf("[ERROR] aclnnTransMatmulWeight failed. ret = %d \n", ret);return ret);
ACL_CHECK(aclrtSynchronizeStreamWithTimeout(args.stream, 20000));
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 launchOneThreadQuantMatmulAllReduce(Args &args) {
int ret;
ret = aclrtSetCurrentContext(args.context);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
char hcom_name[128];
ret = HcclGetCommName(args.hcclComm, hcom_name);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("[ERROR] HcclGetCommName failed. ret = %d \n", ret); return -1);
LOG_PRINT("[INFO] rank %d hcom: %s stream: %p, context : %p\n", args.rankId, hcom_name, args.stream,
args.context);
std::vector<int64_t> x1Shape = {32, 64};
std::vector<int64_t> x2Shape = {64, 128};
std::vector<int64_t> biasShape = {128};
std::vector<int64_t> dequantScaleShape = {128};
std::vector<int64_t> x3Shape = {32, 128};
std::vector<int64_t> outShape = {32, 128};
void *x1DeviceAddr = nullptr;
void *x2DeviceAddr = nullptr;
void *biasDeviceAddr = nullptr;
void *dequantScaleDeviceAddr = nullptr;
void *x3DeviceAddr = nullptr;
void *outDeviceAddr = nullptr;
aclTensor *x1 = nullptr;
aclTensor *x2 = nullptr;
aclTensor *bias = nullptr;
aclTensor *dequantScale = nullptr;
aclTensor *x3 = nullptr;
aclTensor *out = nullptr;
int64_t commTurn = 0;
int64_t streamMode = 1;
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
void *workspaceAddr = nullptr;
long long x1ShapeSize = GetShapeSize(x1Shape);
long long x2ShapeSize = GetShapeSize(x2Shape);
long long biasShapeSize = GetShapeSize(biasShape);
long long dequantScaleShapeSize = GetShapeSize(dequantScaleShape);
long long x3ShapeSize = GetShapeSize(x3Shape);
long long outShapeSize = GetShapeSize(outShape);
std::vector<int8_t> x1HostData(x1ShapeSize, 0);
std::vector<int8_t> x2HostData(x2ShapeSize, 0);
std::vector<int32_t> biasHostData(biasShapeSize, 0);
std::vector<uint64_t> dequantScaleHostData(dequantScaleShapeSize, 0);
std::vector<int16_t> x3HostData(x3ShapeSize, 0);
std::vector<int16_t> outHostData(outShapeSize, 0);
// 创建 tensor
ret = CreateAclTensor(x1HostData, x1Shape, &x1DeviceAddr, aclDataType::ACL_INT8, &x1);
CHECK_RET(ret == ACL_SUCCESS, return ret);
if (args.format == "NZ") {
ret = CreateWeightNzAclTensor(x2HostData, x2Shape, &x2DeviceAddr, aclDataType::ACL_INT8, &x2, args);
} else {
ret = CreateAclTensor(x2HostData, x2Shape, &x2DeviceAddr, aclDataType::ACL_INT8, &x2);
}
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(dequantScaleHostData, dequantScaleShape, &dequantScaleDeviceAddr,
aclDataType::ACL_UINT64, &dequantScale);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(x3HostData, x3Shape, &x3DeviceAddr, aclDataType::ACL_FLOAT16, &x3);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 调用第一段接口
ret = aclnnQuantMatmulAllReduceGetWorkspaceSize(x1, x2, bias, x3, dequantScale, hcom_name, "sum", commTurn, streamMode, out,
&workspaceSize, &executor);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclnnQuantMatmulAllReduceGetWorkspaceSize 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);
}
// 调用第二段接口
ret = aclnnQuantMatmulAllReduce(workspaceAddr, workspaceSize, executor, args.stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnQuantMatmulAllReduce failed. ERROR: %d\n", ret); return ret);
//（固定写法）同步等待任务执行结束
ret = aclrtSynchronizeStreamWithTimeout(args.stream, 10000);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
LOG_PRINT("device%d aclnnQuantMatmulAllReduce execute success \n", args.rankId);
// 释放device资源，需要根据具体API的接口定义修改
if (x1 != nullptr) {
aclDestroyTensor(x1);
}
if (x2 != nullptr) {
aclDestroyTensor(x2);
}
if (bias != nullptr) {
aclDestroyTensor(bias);
}
if (dequantScale != nullptr) {
aclDestroyTensor(dequantScale);
}
if (x3 != nullptr) {
aclDestroyTensor(x3);
}
if (out != nullptr) {
aclDestroyTensor(out);
}
if (x1DeviceAddr != nullptr) {
aclrtFree(x1DeviceAddr);
}
if (x2DeviceAddr != nullptr) {
aclrtFree(x2DeviceAddr);
}
if (biasDeviceAddr != nullptr) {
aclrtFree(biasDeviceAddr);
}
if (dequantScaleDeviceAddr != nullptr) {
aclrtFree(dequantScaleDeviceAddr);
}
if (x3DeviceAddr != nullptr) {
aclrtFree(x3DeviceAddr);
}
if (outDeviceAddr != nullptr) {
aclrtFree(outDeviceAddr);
}
if (workspaceSize > 0) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(args.stream);
HcclCommDestroy(args.hcclComm);
aclrtDestroyContext(args.context);
aclrtResetDevice(args.rankId);
return 0;
}
int main(int argc, char *argv[]) {
int ret;
int32_t devices[ndev];
for (int i = 0; i < ndev; i++) {
devices[i] = i;
}
HcclComm comms[128];
ret = aclInit(nullptr);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
// 初始化集合通信域
for (int i = 0; i < ndev; i++) {
ret = aclrtSetDevice(devices[i]);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
}
ret = HcclCommInitAll(ndev, devices, comms);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("HcclCommInitAll failed. ERROR: %d\n", ret); return ret);
Args args[ndev];
aclrtStream stream[ndev];
aclrtContext context[ndev];
for (uint32_t rankId = 0; rankId < ndev; rankId++) {
ret = aclrtSetDevice(rankId);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
ret = aclrtCreateContext(&context[rankId], rankId);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
ret = aclrtCreateStream(&stream[rankId]);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
}
// 启动多线程
std::vector<std::unique_ptr<std::thread>> threads(ndev);
for (uint32_t rankId = 0; rankId < ndev; rankId++) {
args[rankId].rankId = rankId;
args[rankId].hcclComm = comms[rankId];
args[rankId].stream = stream[rankId];
args[rankId].context = context[rankId];
threads[rankId].reset(
new(std::nothrow) std::thread(&launchOneThreadQuantMatmulAllReduce, std::ref(args[rankId])));
}
for (uint32_t rankId = 0; rankId < ndev; rankId++) {
threads[rankId]->join();
}
aclFinalize();
return 0;
}