1. 概述

通过pybind11，生成autodock-gpu的python binding，方便python直接调用。

https://github.com/ccsb-scripps/AutoDock-GPU

2. autodock-gpu的编译命令

• 编译命令，生成可执行文件

GPU_INCLUDE_PATH=/usr/local/cuda-11.2/include GPU_LIBRARY_PATH=/usr/local/cuda-11.2/lib64 DEVICE=CUDA make DEVICE=CUDA NUMWI=256 OVERLAP=ON TARGETS="86"

会自动生成host/src/performdocking.cpp和host/inc/performdocking.h两个文件，后面的编译需要这两个文件
因为修改了main代码，这两个文件没办法再次生成，所以一并提交到repo中，可能在其他GPU环境中导致so文件不可用，请留意

• 不试用makefile，单独编译生成autodock_gpu_256wi可执行文件

g++ ./host/src/calcenergy.cpp ./host/src/getparameters.cpp ./host/src/main.cpp ./host/src/miscellaneous.cpp ./host/src/performdocking.cpp ./host/src/processgrid.cpp ./host/src/processligand.cpp ./host/src/processresult.cpp ./host/src/setup.cpp \
    -std=c++17 -I./common -I./host/inc -I/usr/local/cuda-11.2/include -I./cuda \
    -L/usr/local/cuda-11.2/lib64 -Wl,-rpath=/usr/local/cuda-11.2/lib64: \
    -DVERSION=\"v1.5.3-73-gf5cf6ffdd0c5b3f113d5cc424fabee51df04da7e\" kernels.o -lcurand -lcudart \
    -o./bin/autodock_gpu_256wi \
    -DGPU_DEVICE -DN256WI -DUSE_PIPELINE -fopenmp -O3

3. pybind11

• 增加main函数的wrap函数，难点是对chat* argv[]的处理，参考

#include <pybind11/pybind11.h>
#include <stdio.h>

#if PY_VERSION_HEX < 0x03000000
#define MyPyText_AsString PyString_AsString
#else
#define MyPyText_AsString PyUnicode_AsUTF8
#endif

namespace py = pybind11;

int run(py::list inlist) {
    int argc = (int)inlist.size();
    char** argv = (char**)malloc(argc * sizeof(char*));

    for (int i = 0; i < argc; ++i)
        argv[i] = (char*)MyPyText_AsString(inlist[i].ptr());

    for (int i = 0; i < argc; ++i)
        fprintf(stderr, "%s\n", argv[i]);

    free(argv);

    return 0;
}

PYBIND11_MODULE(example, m) {
    m.def("run", &run, "runs the example");
}

还有一些其他的修改点，难度不大，这里不详细赘述

• 生成pybind11的so文件

g++ -O3 -Wall -shared -fPIC $(python3 -m pybind11 --includes) -fno-inline -Wl,-Bsymbolic-functions \
     ./host/src/calcenergy.cpp ./host/src/getparameters.cpp ./host/src/main.cpp ./host/src/miscellaneous.cpp ./host/src/performdocking.cpp ./host/src/processgrid.cpp ./host/src/processligand.cpp ./host/src/processresult.cpp ./host/src/setup.cpp \
    -std=c++17 -I./common -I./host/inc -I/usr/local/cuda-11.2/include -I./cuda \
    -L/usr/local/cuda-11.2/lib64 -Wl,-rpath=/usr/local/cuda-11.2/lib64: \
    -DVERSION=\"v1.5.3-73-gf5cf6ffdd0c5b3f113d5cc424fabee51df04da7e\" kernels.o -lcurand -lcudart \
    -o./bin/autodock$(python3-config --extension-suffix) \
    -DGPU_DEVICE -DN256WI -DUSE_PIPELINE -fopenmp

• 测试

在上一步生成的autodock.cpython-310-x86_64-linux-gnu.so同一目录，创建python文件如下，然后执行

import autodock


argv = "--lfile /home/shuzhang/ai/code/moldock/autodock/output/tmpnnuuab_g.pdbqt --ffile /data/autodock/grid/0cb544cb1474ff6d917fe409598886cb/protein.maps.fld --devnum 2 --ngen 1 --nrun 2 --stopstd 1.999"
mols = autodock.run(argv.split())

to_float = lambda x: float(x.strip())
for i, mol in enumerate(mols):
    print(f"----{i}")
    for atom in mol.split(";"):
        if not atom:
            continue
        x,y,z = atom.split(",")
        print(f"{to_float(x):.5f}, {to_float(y):.5f}, {to_float(z):.5f}")

4. 总结

pybind11功能强大，且简单易用，更多文档参考pybind11官方文档。