3DGS代码调试

2025-09-20

1. 调试

先在服务器中安装c/c++插件

1.1 修改文件（跳过）

同时调试python和cuda代码需要做额外配置：

1 2	cd ~/code/gaussian-splatting grep -R "setup.py" -n

根据输出结果发现这三个目录下有setup.py

submodules/simple-knn
submodules/fused-ssim

submodules/diff-gaussian-rasterization

scripts/debug.sh:36:  "$PY" setup.py clean  > build.log 2>&1 || true
scripts/debug.sh:37:  "$PY" setup.py build >> build.log 2>&1
匹配到二进制文件 .git/modules/submodules/simple-knn/index
匹配到二进制文件 .git/modules/submodules/fused-ssim/index
匹配到二进制文件 .git/modules/submodules/diff-gaussian-rasterization/index
test.py:27:            print(f"[WARN] {b} 里找不到 {WANT} 的 build 目录；请在该子模块用当前 python 重新 `python setup.py build`")
submodules/simple-knn/simple_knn.egg-info/SOURCES.txt:3:setup.py
submodules/fused-ssim/fused_ssim.egg-info/SOURCES.txt:4:setup.py
submodules/fused-ssim/build.log:2:  File "setup.py", line 39, in <module>
submodules/fused-ssim/build.log:6:  File "setup.py", line 39, in <module>
submodules/diff-gaussian-rasterization/diff_gaussian_rasterization.egg-info/SOURCES.txt:5:setup.py

修改code/gaussian-splatting/submodules/simple-knn/setup.py:

from setuptools import setup
from torch.utils.cpp_extension import CUDAExtension, BuildExtension
import os

cxx_compiler_flags = []

if os.name == 'nt':
    cxx_compiler_flags.append("/wd4624")

# 判断是否调试
DEBUG = os.getenv("CUDA_DEBUG_BUILD", "0") == "1"
if DEBUG:
    extra = {
        # Linux 下把 -fPIC 也加上，更稳
        "cxx":  (cxx_compiler_flags + (["/Od", "/Z7"] if os.name == 'nt' else ["-O0", "-g", "-fPIC"])),
        "nvcc": (["-O0", "-g", "-G", "-lineinfo"] + ([] if os.name == 'nt' else ["-Xcompiler", "-fPIC"])),
    }
else:
    extra = {"nvcc": [], "cxx": cxx_compiler_flags}


setup(
    name="simple_knn",
    ext_modules=[
        CUDAExtension(
            name="simple_knn._C",
            sources=[
            "spatial.cu", 
            "simple_knn.cu",
            "ext.cpp"],
            extra_compile_args=extra,
            )
        ],
    cmdclass={
        'build_ext': BuildExtension
    }
)

修改code/gaussian-splatting/submodules/fused-ssim/setup.py：

# from setuptools import setup
# from torch.utils.cpp_extension import CUDAExtension, BuildExtension

# setup(
#     name="fused_ssim",
#     packages=['fused_ssim'],
#     ext_modules=[
#         CUDAExtension(
#             name="fused_ssim_cuda",
#             sources=[
#             "ssim.cu",
#             "ext.cpp"])
#         ],
#     cmdclass={
#         'build_ext': BuildExtension
#     }
# )


from setuptools import setup
from torch.utils.cpp_extension import CUDAExtension, BuildExtension



# 判断是否调试
import os
DEBUG = os.getenv("CUDA_DEBUG_BUILD", "0") == "1"
if DEBUG:
    setup(
        name="fused_ssim",
        packages=['fused_ssim'],
        ext_modules=[
            CUDAExtension(
                name="fused_ssim_cuda",
                sources=[
                "ssim.cu",
                "ext.cpp"],
                extra_compile_args={
                    "cxx":  (cxx_compiler_flags + (["/Od", "/Z7"] if os.name == 'nt' else ["-O0", "-g"])),
                    "nvcc": (["-O0", "-g", "-G", "-lineinfo"] + ([] if os.name == 'nt' else ["-Xcompiler", "-fPIC"])),
                }
                )
            ],
        cmdclass={
            'build_ext': BuildExtension
        }
    )

else:
    setup(
        name="fused_ssim",
        packages=['fused_ssim'],
        ext_modules=[
            CUDAExtension(
                name="fused_ssim_cuda",
                sources=[
                "ssim.cu",
                "ext.cpp"])
            ],
        cmdclass={
            'build_ext': BuildExtension
        }
    )

修改code/gaussian-splatting/submodules/diff-gaussian-rasterization/setup.py：

from setuptools import setup
from torch.utils.cpp_extension import CUDAExtension, BuildExtension
import os

this_dir = os.path.dirname(os.path.abspath(__file__))
glm_inc  = os.path.join(this_dir, "third_party/glm/")

is_windows = os.name == "nt"
DEBUG = os.getenv("CUDA_DEBUG_BUILD", "0") == "1"

if DEBUG:
    cxx_flags  = (["/Od", "/Z7"] if is_windows else ["-O0", "-g"])
    nvcc_flags = (["-O0", "-g", "-G", "-lineinfo"] +
                  ([] if is_windows else ["-Xcompiler", "-fPIC"]) +
                  ["-I" + glm_inc])
    extra = {"cxx": cxx_flags, "nvcc": nvcc_flags}
else:
    # Release：保留你原来的 include，不额外加任何 flag
    extra = {"nvcc": ["-I" + glm_inc]}

setup(
    name="diff_gaussian_rasterization",
    packages=["diff_gaussian_rasterization"],
    ext_modules=[
        CUDAExtension(
            name="diff_gaussian_rasterization._C",
            sources=[
                "cuda_rasterizer/rasterizer_impl.cu",
                "cuda_rasterizer/forward.cu",
                "cuda_rasterizer/backward.cu",
                "rasterize_points.cu",
                "ext.cpp",
            ],
            extra_compile_args=extra,
        )
    ],
    cmdclass={"build_ext": BuildExtension},
)

新建并运行：

#!/usr/bin/env bash
set -euo pipefail

# --- 绝对路径：固定用 3dgs 环境的 python/pip/ninja ---
ENV_DIR="$HOME/anaconda3/envs/3dgs"
PY="$ENV_DIR/bin/python"
PIP="$ENV_DIR/bin/pip"
NINJA="$ENV_DIR/bin/ninja"

# --- 能用就静默激活 conda：失败也别吵 ---
if command -v conda >/dev/null 2>&1; then
  source "$(conda info --base)/etc/profile.d/conda.sh" >/dev/null 2>&1 || true
  conda activate 3dgs >/dev/null 2>&1 || true
fi

# --- ninja：用你自己的源安装；已有就跳过 ---
if ! "$NINJA" --version >/dev/null 2>&1; then
  conda install -y -n 3dgs ninja >/dev/null 2>&1 || "$PIP" install -q ninja
fi

# --- 开启调试构建（只 build，不 install） ---
export CUDA_DEBUG_BUILD=1

# 当前 Python ABI（挑选正确的 build/lib.* 目录）
PYTAG=$("$PY" - <<'PY'
import sys; print(f"cpython-{sys.version_info.major}{sys.version_info.minor}")
PY
)

# 小工具：安静构建并返回 build/lib.* 路径（只回路径，日志写文件）
quiet_build() {
  local dir="$1"
  [ -d "$dir" ] || { echo ""; return; }
  cd "$dir"
  # 把构建日志写到 build.log（不污染 stdout）
  "$PY" setup.py clean  > build.log 2>&1 || true
  "$PY" setup.py build >> build.log 2>&1

  # 选与当前 Python 匹配的 lib 目录；没有就取最新的
  local best=""
  for p in "$PWD"/build/lib.*; do
    [ -d "$p" ] || continue
    case "$p" in *"$PYTAG"*) best="$p";; esac
  done
  [ -n "$best" ] || best=$(ls -dt "$PWD"/build/lib.* 2>/dev/null | head -n1 || true)
  echo "$best"
}

ROOT="$HOME/code/gaussian-splatting"

SIMPLE_LIB=$(quiet_build "$ROOT/submodules/simple-knn")


FSSIM_LIB=$(quiet_build "$ROOT/submodules/fused-ssim")


DGR_LIB=$(quiet_build "$ROOT/submodules/diff-gaussian-rasterization")

# 打印各自路径（纯净）
echo "$SIMPLE_LIB"
echo "$FSSIM_LIB"
echo "$DGR_LIB"

# 组合出干净的 PYTHONPATH（去空）
PYTHONPATH_VAL=$(printf "%s\n%s\n%s\n" "$SIMPLE_LIB" "$FSSIM_LIB" "$DGR_LIB" | awk 'NF' | paste -sd: -)

echo
echo "export PYTHONPATH=\"$PYTHONPATH_VAL\""
echo "(复制上面这一行到当前终端，即可临时使用调试版 .so)"
echo "注：编译日志保存在各子模块的 build.log（如 submodules/simple-knn/build.log）"

1.1 修改文件

1.2 配置文件


{
    "version": "0.2.0",
    "configurations": [
        // 1. Python调试配置（和你原有逻辑一致，仅微调路径映射）
        {
            "name": "Python Attach (6789)",
            "type": "debugpy-old",
            "request": "attach",
            "connect": {
                "host": "localhost",
                "port": 6789
            },
            "justMyCode": false,
            // "pathMappings": [
            //     {
            //         "localRoot": "/home/csu_kcc/code/gaussian-splatting", // 本地项目根目录（VSCode打开的文件夹）
            //         "remoteRoot": "/home/csu_kcc/code/gaussian-splatting" // 服务器上项目的绝对路径（必须写死，避免变量出错）
            //     }
            // ],
            // "stopOnEntry": true // 附加后不自动暂停，方便后续执行到CUDA调用前
        },
        // 2. CUDA调试配置（关键修正）
        {
            "name": "CUDA Attach to Python",
            "type": "cppdbg", // 核心修正：VSCode通过cppdbg类型调用cuda-gdb
            "request": "attach",
            "MIMode": "gdb",
            // 修正：sudo调用cuda-gdb，确保权限（依赖免密码配置）
            "miDebuggerPath": "/usr/local/cuda/bin/cuda-gdb",
            "processId": "${command:pickProcess}", // 选择Python进程（和之前一样）
            "program": "/home/csu_kcc/anaconda3/envs/3dgs/bin/python", // 你的Python解释器（必须绝对路径）
            "environment": [
                {
                    "name": "PYTHONPATH",
                    "value": "/home/csu_kcc/code/gaussian-splatting/submodules/simple-knn/build/lib.linux-x86_64-cpython-37:/home/csu_kcc/code/gaussian-splatting/submodules/fused-ssim/build/lib.linux-x86_64-cpython-37:/home/csu_kcc/code/gaussian-splatting/submodules/diff-gaussian-rasterization/build/lib.linux-x86_64-cpython-37"
                }
            ],
            "setupCommands": [
                {
                    "text": "handle SIGURG nostop noprint pass",
                    "ignoreFailures": true
                },
                // 启用符号美化（方便查看变量）
                { "text": "-enable-pretty-printing", "ignoreFailures": true },
                // 禁止CUDA核函数启动时自动断点（避免一启动就卡）
                { "text": "set cuda break_on_launch none" },
                // 加载CUDA调试符号（避免“无调试符号”警告）
                { "text": "set debug-file-directory /usr/local/cuda/lib/debug" }
            ],
            // 核心修正：用集成终端处理权限（若免密码失效，可在VSCode内输入密码）
            "externalConsole": false,
            "console": "integratedTerminal",
            // 确保调试器能找到CUDA代码（映射.cu文件路径）
            "sourceFileMap": {
                "/home/csu_kcc/code/gaussian-splatting": "/home/csu_kcc/code/gaussian-splatting"
            }
        }
    ]
}

先运行代码 python -m debugpy --listen localhost:6789 --wait-for-client train.py -s ~/code/gaussian-splatting/data/input/tandt_db/truck -m ~/code/gaussian-splatting/data/output/tandt_db/truck --eval
然后在左侧调试栏，选择python attach，点击调试按钮
当调试到需要进入cuda的部分，在调试栏选择CUDA Attach to Python，然后选择刚刚运行的进程（搜索train.py）
选择之后显示：此操作可能会危害计算机需要管理员权限，输入y

如果输入y之后直接显示如下内容，则转1.3

Multiple identities can be used for authentication:
 1.  user,,, (user)
 2.  xxx
Choose identity to authenticate as (1-2): [1] + Stopped (tty input)        /usr/bin/pkexec "/usr/bin/sudo" --interpreter=mi --tty=${DbgTerm} 0<"/tmp/Microsoft-MIEngine-In-ejoxxylu.2hh" 1>"/tmp/Microsoft-MIEngine-Out-rv5zyxew.go0"
You have stopped jobs.

1.3 cuda调试配置

修改配置文件

# 在服务器上执行，允许vs code调试器无需密码使用sudo
echo "$USER ALL=(ALL) NOPASSWD: /usr/bin/gdb, /usr/local/cuda-11.6/bin/cuda-gdb" | sudo tee /etc/sudoers.d/vscode-debug
sudo chmod 0440 /etc/sudoers.d/vscode-debug
sudo visudo -c -f /etc/sudoers.d/vscode-debug

若后续不需要调试，直接删除该文件即可完全撤销设置：
1
sudo rm /etc/sudoers.d/vscode-debug
输入：如果路径不同，需要修改sudoers.d/vscode-debug文件中的路径（修改后记得重新用visudo -c验证）。

1	which cuda-gdb # 输出应与配置中的路径一致

配置之后如果调试输入y之后仍然直接停止，则：

1	sudo nano /etc/polkit-1/localauthority/50-local.d/99-vscode-debug.pkla

写入：

[Allow Debuggers without Password]
Identity=unix-user:<你的用户名>
Action=org.freedesktop.policykit.exec
ResultAny=yes
ResultInactive=yes
ResultActive=yes

然后

sudo chmod 644 /etc/polkit-1/localauthority/50-local.d/99-vscode-debug.pkla
# 重启PolicyKit守护进程（不同系统可能有差异，尝试以下命令）
sudo systemctl restart polkit
# 若上面命令失败，尝试：
sudo service polkit restart

1.4 cuda调试

输入y之后进入调试，发现cuda调试进程有一个进程显示exception：

1	unable toretrive stack trace. the message is improperly formatted or was damaged in transit

查看CUDA 调试符号路径

1 2	ls /usr/local/cuda-11.6/lib64/debug # 或者ls /usr/local/cuda-11.6/lib/debug

发现里面不存在debug文件夹，安装cuda调试符号包

1	sudo apt install cuda-dbg-11-6

如果发现 E: 无法定位软件包 cuda-dbg-11-6

# 下载 CUDA 11.6 的源配置包（Ubuntu 20.04 对应代号 focal）
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
# 将源优先级文件复制到系统目录
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
# 下载并安装 CUDA 11.6 源（注意：版本号必须严格匹配 11.6）
wget https://developer.download.nvidia.com/compute/cuda/11.6.0/local_installers/cuda-repo-ubuntu2004-11-6-local_11.6.0-510.39.01-1_amd64.deb
# 安装源包（会自动添加 NVIDIA 签名密钥）
sudo dpkg -i cuda-repo-ubuntu2004-11-6-local_11.6.0-510.39.01-1_amd64.deb
# 更新 apt 缓存（让系统识别新添加的源）
cd /var/cuda-repo-ubuntu2004-11-6-local/
sudo apt-key add 7fa2af80.pub
sudo apt update
sudo apt install cuda-dbg-11-6