在MLU-OPS™ 开发算子时，如果需要用到某个基础功能，其功能已经由CNNL 算子实现，我们不妨直接在MLU-OPS™ 代码中调用CNNL 的API，快速达成算子开发的目的。下面3个示例介绍了如何在 MLU-OPS™ 代码中实现调用CNNL 算子的方法。
在算子中调用CNNL tranpose 算子。
// 需要包含该文件
#include "utils/cnnl_helper.h"
// https://github.com/Cambricon/mlu-ops/blob/master/kernels/three_nn_forward/three_nn_forward.cpp
static mluOpStatus_t transposeTensor(
const mluOpHandle_t handle, const mluOpTensorDescriptor_t input_desc,
const void *input, const int *permute,
const mluOpTensorDescriptor_t workspace_dst_desc, void *workspace_dst,
void *transpose_workspace, size_t transpose_workspace_size) {
const int input_dim = input_desc->dim;
// step1. 将mluOpHandle、mluOpTensorDescriptor_t 等参数转换为 CNNL 公共参数类型
// 宏 DEFINE_CREATE_AND_SET_CNNL_HANDLE 内部定义并创建 CNNL handle，
// 并将 MLU-OPS handle 信息转换设置到 CNNL handle
DEFINE_CREATE_AND_SET_CNNL_HANDLE(handle, cnnl_handle);
// 宏 DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR 内部定义并创建 CNNL TensorDescriptor
// (区别于各算子特有的descriptor，例如上述的 cnnlTranposeDescriptor，通用的TensorDescriptor通过宏操作更方便)，
// 并将 MLU-OPS TensorDescriptor 信息转换设置到 CNNL TensorDescriptor
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(input_desc, cnnl_input_desc);
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(workspace_dst_desc,
cnnl_workspace_dst_desc);
// step2. 定义 cnnlTranspose_v2 cnnlTranposeDescriptor 类型入参
// MLU-OPS代码中直接定义cnnl cnnl_trans_desc 变量
cnnlTransposeDescriptor_t cnnl_trans_desc = NULL;
// 直接调用 CNNL api 创建 cnnlTranposeDescriptor
CALL_CNNL(cnnlCreateTransposeDescriptor(&cnnl_trans_desc));
// 直接调用 CNNL api 设置 cnnlTranposeDescriptor
CALL_CNNL(cnnlSetTransposeDescriptor(cnnl_trans_desc, input_dim, permute));
// step3. 调用 cnnlTranspose_v2
// 直接调用 CNNL 算子 api
CALL_CNNL(cnnlTranspose_v2(cnnl_handle, cnnl_trans_desc, cnnl_input_desc,
input, cnnl_workspace_dst_desc, workspace_dst,
transpose_workspace, transpose_workspace_size));
// step4. Destroy cnnlTranspose_v2 cnnlTranposeDescriptor 类型入参
// 直接调用 CNNL api 销毁 cnnlTranposeDescriptor
CALL_CNNL(cnnlDestroyTransposeDescriptor(cnnl_trans_desc));
//step5. Destroy CNNL 公共类型参数 handle 及 tensor descriptor等.
// 宏 DESTROY_CNNL_TENSOR_DESCRIPTOR 内部销毁新建的 CNNL TensorDescriptor
DESTROY_CNNL_TENSOR_DESCRIPTOR(cnnl_workspace_dst_desc);
DESTROY_CNNL_TENSOR_DESCRIPTOR(cnnl_input_desc);
// 宏 DESTROY_CNNL_TENSOR_DESCRIPTOR 内部销毁新建的 CNNL handle
DESTROY_CNNL_HANDLE(cnnl_handle);
return MLUOP_STATUS_SUCCESS;
}
在算子中调用CNNL addN 算子的GetWorkspaceSize, 与示例1不同的点在于，当输入的tensor是数组时的处理方法。
// 需要包含该文件
#include "utils/cnnl_helper.h"
mluOpStatus_t GetAddNWorkspaceSize(
mluOpHandle_t handle, const mluOpTensorDescriptor_t input_descs[],
const uint32_t input_num, const mluOpTensorDescriptor_t output_desc,
size_t *workspace_size) {);
// 当对应的输入是 MLU-OPS TensorDescriptor 数组变量时
// 可先定义并malloc 同样大小的CNNL TensorDescriptor数组变量
cnnlTensorDescriptor_t *_input_descs = (cnnlTensorDescriptor_t *)malloc(
sizeof(cnnlTensorDescriptor_t) * input_num);
// 宏 CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR 内部创建 CNNL TensorDescriptor
// (不需要再定义)，通过循环方式并将每个MLU-OPS TensorDescriptor 信息转换设置到对应的每个
// CNNL TensorDescriptor
for (int i = 0; i < input_num; i++) {
CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(input_descs[i], _input_descs[i]);
}
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(output_desc, _output_desc);
CALL_CNNL(cnnlGetAddNWorkspaceSize(_handle, _input_descs, input_num,
_output_desc, workspace_size);
// 循环销毁每个 CNNL TensorDescriptor
for (int i = 0; i < input_num; i++) {
DESTROY_CNNL_TENSOR_DESCRIPTOR(_input_descs[i]);
}
// 释放malloc的空间
free(_input_descs);
DESTROY_CNNL_TENSOR_DESCRIPTOR(_output_desc);
DESTROY_CNNL_HANDLE(_handle);
return MLUOP_STATUS_SUCCESS;
}
在算子中调用CNNL matmul 算子的相关接口。
// https://github.com/Cambricon/mlu-ops/blob/master/kernels/indice_convolution_backward_data/indice_convolution_backward_data.cpp
// 需要包含该文件
#include "utils/cnnl_helper.h"
mluOpStatus_t MLUOP_WIN_API mluOpGetIndiceConvolutionBackwardDataWorkspaceSize(
mluOpHandle_t handle, const mluOpTensorDescriptor_t output_grad_desc,
const mluOpTensorDescriptor_t filters_desc,
const mluOpTensorDescriptor_t indice_pairs_desc,
const mluOpTensorDescriptor_t input_grad_desc, const int64_t indice_num[],
const int64_t inverse, size_t *workspace_size) {
......
// matmul workspace
{
mluOpTensorDescriptor_t sub_filters_desc;
mluOpTensorDescriptor_t output_grad_condence_desc;
mluOpTensorDescriptor_t input_grad_condence_desc;
// 在 MLU-OPS 代码中直接定义 CNNL 变量
cnnlMatMulDescriptor_t cnnl_matmul_desc;
cnnlMatMulHeuristicResult_t cnnl_heuristic_result;
cnnlMatMulAlgo_t cnnl_matmul_algo;
MLUOP_CHECK(mluOpCreateTensorDescriptor(&sub_filters_desc));
int sub_filter_dims[2] = {(int)(dxc), (int)(dyc)};
MLUOP_CHECK(mluOpSetTensorDescriptor(sub_filters_desc, MLUOP_LAYOUT_ARRAY,
filters_desc->dtype, 2,
sub_filter_dims));
int is_trans_a = 0, is_trans_b = 1;
int tf32_flag_int = 0;
// 直接调用 CNNL api 对结构进行创建、设置
CALL_CNNL(cnnlMatMulDescCreate(&cnnl_matmul_desc));
CALL_CNNL(cnnlSetMatMulDescAttr(cnnl_matmul_desc, CNNL_MATMUL_DESC_TRANSA,
&(is_trans_a), sizeof(is_trans_a)));
CALL_CNNL(cnnlSetMatMulDescAttr(cnnl_matmul_desc, CNNL_MATMUL_DESC_TRANSB,
&(is_trans_b), sizeof(is_trans_b)));
CALL_CNNL(cnnlSetMatMulDescAttr(cnnl_matmul_desc, CNNL_MATMUL_ALLOW_TF32,
&(tf32_flag_int), sizeof(tf32_flag_int)));
MLUOP_CHECK(mluOpCreateTensorDescriptor(&output_grad_condence_desc));
int output_grad_condence_dims[2] = {(int)(max_indice_num), (int)(dyc)};
MLUOP_CHECK(mluOpSetTensorDescriptor(
output_grad_condence_desc, MLUOP_LAYOUT_ARRAY, output_grad_desc->dtype,
2, output_grad_condence_dims));
MLUOP_CHECK(mluOpCreateTensorDescriptor(&input_grad_condence_desc));
int input_grad_condence_dims[2] = {(int)(max_indice_num), (int)(dxc)};
MLUOP_CHECK(mluOpSetTensorDescriptor(
input_grad_condence_desc, MLUOP_LAYOUT_ARRAY, input_grad_desc->dtype, 2,
input_grad_condence_dims));
// 通过宏内部定义并创建 CNNL Handle，TensorDescritpor，
// 并将 MLU-OPS Handle，TensorDescritpor 信息转换设置到对应
// CNNL Handle，TensorDescritpor
DEFINE_CREATE_AND_SET_CNNL_HANDLE(handle, cnnl_handle);
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(sub_filters_desc,
cnnl_sub_filters_desc);
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(output_grad_condence_desc,
cnnl_output_grad_condence_desc);
DEFINE_CREATE_AND_SET_CNNL_TENSOR_DESCRIPTOR(input_grad_condence_desc,
cnnl_input_grad_condence_desc);
// 直接调用 CNNL api 对结构进行创建、设置
CALL_CNNL(cnnlCreateMatMulHeuristicResult(&cnnl_heuristic_result));
CALL_CNNL(cnnlMatMulAlgoCreate(&cnnl_matmul_algo));
// set matmul heuristic_result & algorithm
int requested_algo_count = 1, return_algo_count = 0;
// 直接调用 CNNL api
CALL_CNNL(cnnlGetMatMulAlgoHeuristic(
cnnl_handle, cnnl_matmul_desc, cnnl_output_grad_condence_desc,
cnnl_sub_filters_desc, cnnl_input_grad_condence_desc,
cnnl_input_grad_condence_desc, NULL, requested_algo_count,
&cnnl_heuristic_result, &return_algo_count));
// launch matmul
size_t workspace_size_matmul = 0;
float alpha_gemm = 1.0f, beta_gemm = 0.0f;
// 直接调用 CNNL api
CALL_CNNL(cnnlGetMatMulHeuristicResult(
cnnl_heuristic_result, cnnl_matmul_algo, &workspace_size_matmul));
// destroy descriptors
// 直接调用 CNNL api 对非宏生成的CNNL 变量销毁
CALL_CNNL(cnnlDestroyMatMulHeuristicResult(cnnl_heuristic_result));
CALL_CNNL(cnnlMatMulDescDestroy(cnnl_matmul_desc));
CALL_CNNL(cnnlMatMulAlgoDestroy(cnnl_matmul_algo));
// 销毁由宏内部定义并创建的变量
DESTROY_CNNL_TENSOR_DESCRIPTOR(cnnl_output_grad_condence_desc);
DESTROY_CNNL_TENSOR_DESCRIPTOR(cnnl_sub_filters_desc);
DESTROY_CNNL_TENSOR_DESCRIPTOR(cnnl_input_grad_condence_desc);
DESTROY_CNNL_HANDLE(cnnl_handle);
MLUOP_CHECK(mluOpDestroyTensorDescriptor(output_grad_condence_desc));
MLUOP_CHECK(mluOpDestroyTensorDescriptor(sub_filters_desc));
MLUOP_CHECK(mluOpDestroyTensorDescriptor(input_grad_condence_desc));
matmul_workspace_size = (uint64_t)workspace_size_matmul;
}
......
return MLUOP_STATUS_SUCCESS;
}