光合开发者社区-论坛

运行时

提问帖

有回复

新人贴

调用thrust::device_vector的resize方法直接崩溃

2026-01-12 10:19:00

我现在在一个有8块海光BW500SM（gfx936）加速卡的docker中测试程序，docker安装的操作系统是Ubuntu22.04，我的程序是在我本机的Ubuntu22.04docker中用cmake编译的，远程传输到海光BW500SM卡的docker中运行，调用thrust::device_vector的resize方法分配200KB全局内存直接崩溃。为了定位问题，我写了一个测试程序，用hipcc编译运行都正常，但用cmake编译则还是在resize时崩溃。我的cmake除了有一个下面写的结构与HIP有关，其他都是与HIP无关的东西，请问有谁碰上过与我类似的问题吗？怎么解决的？ add_compile_options( -arch=gfx906 -arch=gfx926 -arch=gfx928 -arch=gfx936 #-rdc=true # 启用 RDC #这行一开始是开启的，但resize时还是崩溃，我注释掉了就没改回来 )

下面是我的测试程序：

#include <hip/hip_runtime.h>

#include <thrust/version.h>

#include <thrust/device_vector.h>

#include <cstdint>

#include <iostream>

struct ss {

uint64_t m;

uint64_t n;

};

int main() {

const size_t size_in_bytes = 200 * 1024; // 200 KB

std::cout << "Thrust version: " << THRUST_VERSION << std::endl;

try {

hipStream_t stream;

hipError_t err = hipStreamCreateWithFlags(&stream, hipStreamNonBlocking);

if (err != hipSuccess) {

std::cerr << "Error: " << std::string(hipGetErrorString(err)) << std::endl;

return -1;

}

// 创建一个空的 device_vector

thrust::device_vector<ss> d_vec;

// 分配 200 KB 内存（即 200 * 1024 个 uint8_t 元素）

d_vec.resize(size_in_bytes);

std::cout << "Successfully allocated " << size_in_bytes << " bytes on GPU." << std::endl;

std::cout << "Device vector size: " << d_vec.size() << " elements." << std::endl;

// 可选：验证分配是否成功（例如通过访问第一个和最后一个元素）

if (!d_vec.empty()) {

thrust::for_each(

thrust::hip::par.on(stream),

d_vec.begin(),

d_vec.end(),

[=] __device__ (ss &s) {

s.m = 1;

s.n = 2;

}

);

std::cout << "Test write successful." << std::endl;

}

if (stream) {

err = hipStreamDestroy(stream);

if (err != hipSuccess) {

std::cerr << "Error: " << std::string(hipGetErrorString(err)) << std::endl;

return -1;

}

} catch (const std::exception& e) {

std::cerr << "Error: " << e.what() << std::endl;

return -1;

}

return 0;

}

下面海光技术支持提供的cmake文件：

cmake_minimum_required(VERSION 3.16)

# Find hipcc executable

find_program(HIPCC_EXECUTABLE hipcc

PATHS /opt/dtk-25.04.3/bin

NO_DEFAULT_PATH

)

if(NOT HIPCC_EXECUTABLE)

find_program(HIPCC_EXECUTABLE hipcc REQUIRED)

endif()

# Find ROCm path

find_path(ROCM_ROOT

NAMES include/hip/hip_runtime.h

PATHS /opt/dtk-25.04.3

DOC "ROCm installation directory"

)

# Set hipcc as the C++ compiler

set(CMAKE_CXX_COMPILER ${HIPCC_EXECUTABLE})

project(resize_test LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 17) # Changed to match your original code

set(CMAKE_CXX_STANDARD_REQUIRED ON)

add_executable(resize-test resize-test.cpp)

# Set HIP-specific compile options with ROCm path and specific architecture

set_property(TARGET resize-test APPEND_STRING PROPERTY

COMPILE_FLAGS " --rocm-path=${ROCM_ROOT} --offload-arch=gfx936")

# Only include architectures that exist in your ROCm installation

# Adjust the architecture based on what's available in your system

target_compile_definitions(resize-test PRIVATE

__HIP_PLATFORM_AMD__=1

THRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP

THRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP

)

这个cmake文件是否可用我还没有测试，估计可用。下面是原来用hipcc直接编译的命令，这个编译运行都没问题：
hipcc -std=c++14 -o ./test ./resize_test.cpp