Files
FastDeploy/fastdeploy/pybind/main.h
Zheng_Bicheng 4ffcfbe726 [Backend] Add RKNPU2 backend support (#456)
* 10-29/14:05
* 新增cmake
* 新增rknpu2 backend

* 10-29/14:43
* Runtime fd_type新增RKNPU代码

* 10-29/15:02
* 新增ppseg RKNPU2推理代码

* 10-29/15:46
* 新增ppseg RKNPU2 cpp example代码

* 10-29/15:51
* 新增README文档

* 10-29/15:51
* 按照要求修改部分注释以及变量名称

* 10-29/15:51
* 修复重命名之后,cc文件中的部分代码还用旧函数名的bug

* 10-29/22:32
* str(Device::NPU)将输出NPU而不是UNKOWN
* 修改runtime文件中的注释格式
* 新增Building Summary ENABLE_RKNPU2_BACKEND输出
* pybind新增支持rknpu2
* 新增python编译选项
* 新增PPSeg Python代码
* 新增以及更新各种文档

* 10-30/14:11
* 尝试修复编译cuda时产生的错误

* 10-30/19:27
* 修改CpuName和CoreMask层级
* 修改ppseg rknn推理层级
* 图片将移动到网络进行下载

* 10-30/19:39
* 更新文档

* 10-30/19:39
* 更新文档
* 更新ppseg rknpu2 example中的函数命名方式
* 更新ppseg rknpu2 example为一个cc文件
* 修复disable_normalize_and_permute部分的逻辑错误
* 移除rknpu2初始化时的无用参数

* 10-30/19:39
* 尝试重置python代码

* 10-30/10:16
* rknpu2_config.h文件不再包含rknn_api头文件防止出现导入错误的问题

* 10-31/14:31
* 修改pybind,支持最新的rknpu2 backends
* 再次支持ppseg python推理
* 移动cpuname 和 coremask的层级

* 10-31/15:35
* 尝试修复rknpu2导入错误

* 10-31/19:00
* 新增RKNPU2模型导出代码以及其对应的文档
* 更新大量文档错误

* 10-31/19:00
* 现在编译完fastdeploy仓库后无需重新设置RKNN2_TARGET_SOC

* 10-31/19:26
* 修改部分错误文档

* 10-31/19:26
* 修复错误删除的部分
* 修复各种错误文档
* 修复FastDeploy.cmake在设置RKNN2_TARGET_SOC错误时,提示错误的信息
* 修复rknpu2_backend.cc中存在的中文注释

* 10-31/20:45
* 删除无用的注释

* 10-31/20:45
* 按照要求修改Device::NPU为Device::RKNPU,硬件将共用valid_hardware_backends
* 删除无用注释以及debug代码

* 11-01/09:45
* 更新变量命名方式

* 11-01/10:16
* 修改部分文档,修改函数命名方式

Co-authored-by: Jason <jiangjiajun@baidu.com>
2022-11-01 11:14:05 +08:00

135 lines
4.2 KiB
C++

// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <pybind11/numpy.h>
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <type_traits>
#include "fastdeploy/runtime.h"
#ifdef ENABLE_VISION
#include "fastdeploy/vision.h"
#include "fastdeploy/pipeline.h"
#endif
#ifdef ENABLE_TEXT
#include "fastdeploy/text.h"
#endif
#include "fastdeploy/core/float16.h"
namespace fastdeploy {
void BindBackend(pybind11::module&);
void BindVision(pybind11::module&);
void BindText(pybind11::module& m);
void BindPipeline(pybind11::module& m);
void BindRKNPU2Config(pybind11::module&);
pybind11::dtype FDDataTypeToNumpyDataType(const FDDataType& fd_dtype);
FDDataType NumpyDataTypeToFDDataType(const pybind11::dtype& np_dtype);
void PyArrayToTensor(pybind11::array& pyarray, FDTensor* tensor,
bool share_buffer = false);
void PyArrayToTensorList(std::vector<pybind11::array>& pyarray,
std::vector<FDTensor>* tensor,
bool share_buffer = false);
pybind11::array TensorToPyArray(const FDTensor& tensor);
#ifdef ENABLE_VISION
cv::Mat PyArrayToCvMat(pybind11::array& pyarray);
#endif
template <typename T>
FDDataType CTypeToFDDataType() {
if (std::is_same<T, int32_t>::value) {
return FDDataType::INT32;
} else if (std::is_same<T, int64_t>::value) {
return FDDataType::INT64;
} else if (std::is_same<T, float>::value) {
return FDDataType::FP32;
} else if (std::is_same<T, double>::value) {
return FDDataType::FP64;
}
FDASSERT(false,
"CTypeToFDDataType only support int32/int64/float32/float64 now.");
return FDDataType::FP32;
}
template <typename T>
std::vector<pybind11::array> PyBackendInfer(
T& self, const std::vector<std::string>& names,
std::vector<pybind11::array>& data) {
std::vector<FDTensor> inputs(data.size());
for (size_t i = 0; i < data.size(); ++i) {
// TODO(jiangjiajun) here is considered to use user memory directly
auto dtype = NumpyDataTypeToFDDataType(data[i].dtype());
std::vector<int64_t> data_shape;
data_shape.insert(data_shape.begin(), data[i].shape(),
data[i].shape() + data[i].ndim());
inputs[i].Resize(data_shape, dtype);
memcpy(inputs[i].MutableData(), data[i].mutable_data(), data[i].nbytes());
inputs[i].name = names[i];
}
std::vector<FDTensor> outputs(self.NumOutputs());
self.Infer(inputs, &outputs);
std::vector<pybind11::array> results;
results.reserve(outputs.size());
for (size_t i = 0; i < outputs.size(); ++i) {
auto numpy_dtype = FDDataTypeToNumpyDataType(outputs[i].dtype);
results.emplace_back(pybind11::array(numpy_dtype, outputs[i].shape));
memcpy(results[i].mutable_data(), outputs[i].Data(),
outputs[i].Numel() * FDDataTypeSize(outputs[i].dtype));
}
return results;
}
} // namespace fastdeploy
namespace pybind11 {
namespace detail {
// Note: use same enum number of float16 in numpy.
// import numpy as np
// print np.dtype(np.float16).num # 23
constexpr int NPY_FLOAT16_ = 23;
// Note: Since float16 is not a builtin type in C++, we register
// fastdeploy::float16 as numpy.float16.
// Ref: https://github.com/pybind/pybind11/issues/1776
template <>
struct npy_format_descriptor<fastdeploy::float16> {
static pybind11::dtype dtype() {
handle ptr = npy_api::get().PyArray_DescrFromType_(NPY_FLOAT16_);
return reinterpret_borrow<pybind11::dtype>(ptr);
}
static std::string format() {
// Note: "e" represents float16.
// Details at:
// https://docs.python.org/3/library/struct.html#format-characters.
return "e";
}
static constexpr auto name = _("float16");
};
} // namespace detail
} // namespace pybind11