[Bug] TVM MIPS32 .so Model Load Failure – ONNX Model Converted to MIPS32 Fails to Load with TVM Runtime

[Leo5050xvjf]

Expected behavior

I expect model.so (a TVM-compiled model for MIPS32) to be successfully loaded using dlopen or TVMModLoadFromFile on my MIPS32 development board, given that libtvm_runtime.so loads and runs correctly

Actual behavior

• libtvm_runtime.so (v0.18.0) loads successfully on the MIPS32 board and executes basic runtime tests.
• However, model.so fails to load with dlopen:

  Failed to load: /path/to/model.so: cannot open shared object file: No such file or directory
  

• Same error occurs with TVMModLoadFromFile, even though the file exists at that path and has correct permissions.

Environment

Operating System:

• Host: Ubuntu 20.04, x86_64
• Target: MIPS32-based development board

TVM Version: v0.18.0

Compiler:

• Host: GCC (x86_64)
• Target: MIPS cross-compiler (mips-linux-gnu-gcc, GCC 7.2.0, glibc 2.29)

TVM Build Configuration:

x86_64 (Host)

#build on x86 PC
cd /path/to/tvm
git checkout v0.18.0
mkdir build-x86_64 && cd build-x86_64
cmake -DCMAKE_BUILD_TYPE=Release -DUSE_LLVM=ON ..
make -j4

• Output: libtvm.so

MIPS32 (Target)

mips32el-toolchain.cmake.cmake

set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_PROCESSOR mipsel)
set(CMAKE_C_COMPILER /opt/mips-gcc720-glibc229/bin/mips-linux-gnu-gcc)
set(CMAKE_CXX_COMPILER /opt/mips-gcc720-glibc229/bin/mips-linux-gnu-g++)
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)

#build on x86 PC
cd /path/to/tvm
git checkout v0.18.0
mkdir build-mips32 && cd build-mips32
cmake \
  -DCMAKE_TOOLCHAIN_FILE=../mips32el-toolchain.cmake \
  -DCMAKE_BUILD_TYPE=Release \
  -DUSE_LLVM=OFF \
  -DUSE_RPC=OFF \
  -DUSE_GRAPH_EXECUTOR=OFF \
  -DUSE_PROFILER=OFF \
  -DUSE_LIBBACKTRACE=OFF \
  ..
make -j4

• Output: libtvm_runtime.so

---

Cross-Compilation for Model

Python script (generate_onnx.py)

import torch
import torch.nn as nn
import torch.onnx


class SimpleCNN(nn.Module):
    def __init__(self):
        super(SimpleCNN, self).__init__()
        self.conv1 = nn.Conv2d(1, 16, 3)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(16, 32, 3)
        self.fc1 = nn.Linear(32 * 5 * 5, 10)

    def forward(self, x):
        x = self.pool(torch.relu(self.conv1(x)))
        x = self.pool(torch.relu(self.conv2(x)))
        x = x.view(-1, 32 * 5 * 5)  
        x = self.fc1(x)
        return x


model = SimpleCNN()
model.eval()

dummy_input = torch.randn(1, 1, 28, 28)


torch.onnx.export(model, dummy_input, "model.onnx", 
                  input_names=["input"], output_names=["output"], 
                  opset_version=11)

print("ONNX model exported as 'model.onnx'")

Python script (generate_so.py)

import tvm
from tvm import relay
import onnx
from tvm.contrib import graph_executor
import tvm.contrib.cc as cc

onnx_model = onnx.load("model.onnx")
input_shape = (1, 1, 28, 28)
mod, params = relay.frontend.from_onnx(onnx_model, {"input": input_shape})

target = "llvm -mtriple=mipsel-linux-gnu -mcpu=mips32r2"

with tvm.transform.PassContext(opt_level=3):
    lib = relay.build(mod, target=target, params=params)

cross_compiler = cc.cross_compiler(
    "/path/to/mips-gcc720-glibc229/bin/mips-linux-gnu-gcc",
    options=["-mfp32", "-mnan=legacy"]
)
lib.export_library("model.so", fcompile=cross_compiler)

• Flags: -mfp32 -mnan=legacy

---

Steps to reproduce

C Test: test_dlopen.c

#include <dlfcn.h>
#include <stdio.h>

int main() {
    void* handle = dlopen("/path/to/model.so", RTLD_LAZY);
    if (!handle) {
        printf("Failed to load: %s\n", dlerror());
        return -1;
    }
    printf("Loaded successfully\n");
    dlclose(handle);
    return 0;
}

• Compile:

/path/to/mips-gcc720-glibc229/bin/mips-linux-gnu-gcc -o test_dlopen test_dlopen.c -ldl

---

C++ Test: test_tvm_runtime.cpp

#include <tvm/runtime/c_runtime_api.h>
#include <iostream>

int main() {
    std::cout << "TVM Runtime Version: " << TVM_VERSION << std::endl;
    TVMAPISetLastError("Test error message");
    const char* error = TVMGetLastError();
    std::cout << "Last Error: " << (error ? error : "None") << std::endl;
    std::cout << "Test completed successfully!" << std::endl;
    return 0;
}

• Compile:

/path/to/mips-gcc720-glibc229/bin/mips-linux-gnu-g++ -o test_tvm_runtime test_tvm_runtime.cpp \
    -I /path/to/tvm/include \
    -I /path/to/tvm/3rdparty/dlpack/include \
    -I /path/to/tvm/3rdparty/dmlc-core/include \
    -L /path/to/tvm/build-mips32 \
    -ltvm_runtime -pthread -std=c++17

---

Deploy and Run on MIPS32

cd /path/to/XXX
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/XXX
chmod +x test_dlopen test_tvm_runtime

./test_tvm_runtime   # Works fine
./test_dlopen        # Fails

---

Output

test_tvm_runtime:
TVM Runtime Version: 0.18.0
Last Error: Test error message
Test completed successfully!

test_dlopen:
Failed to load: /path/to/XXX/model.so: cannot open shared object file: No such file or directory

More Info

file model.so 
model.so: ELF 32-bit LSB shared object, MIPS, MIPS32 rel2 version 1 (SYSV), dynamically linked, with debug_info, not stripped
file libtvm_runtime.so
libtvm_runtime.so: ELF 32-bit LSB shared object, MIPS, MIPS32 rel2 version 1 (SYSV), dynamically linked, with debug_info, not stripped
file libtvm.so
libtvm.so: ELF 64-bit LSB shared object, x86-64, version 1 (GNU/Linux), dynamically linked, BuildID[sha1]=ac5555851eefd24f951e85366f8fd0a5c95987d3, not stripped

If running on a PC x86 platform, the model.so can be successfully loaded and executed.

Triage

compilation

target:mips

component:runtime

area:relay

bug