1. 安装CUDA Toolkit 11.8
从MZ小师妹的摸索过程来看,其他版本的会有bug,12.0的版本太高,11.5的太低(感谢小师妹让我少走弯路)
参考网址:CUDA Toolkit 11.8 Downloads | NVIDIA Developer
在命令行输入命令:
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
sudo sh cuda_11.8.0_520.61.05_linux.run
- 确定自己用的是cuda 11.8:
如果不是,在自己的~/.bashrc文件中添加路径:
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64
export PATH=$PATH:/usr/local/cuda-11.8/bin
-
安装pycuda:
conda install -c conda-forge pycuda -
测试pycuda:
来源 PyCUDA - 上海交大超算平台用户手册 Documentation
python
import pycuda.driver as drv
import pycuda.autoinit
from pycuda.compiler import SourceModule
import numpy
# 定义核函数
mod = SourceModule(
"""
__global__ void add_vectors(float *a, float *b, float *c, int n)
{
int idx = threadIdx.x + blockIdx.x * blockDim.x;
if (idx < n)
{
c[idx] = a[idx] + b[idx];
}
}
"""
)
# 定义向量大小
n = 10000
# 生成随机向量数据
a = numpy.random.randn(n).astype(numpy.float32)
b = numpy.random.randn(n).astype(numpy.float32)
# 分配输出内存空间
c = numpy.zeros_like(a)
# 将输入输出数据复制到 GPU
a_gpu = drv.mem_alloc(a.nbytes)
b_gpu = drv.mem_alloc(b.nbytes)
c_gpu = drv.mem_alloc(c.nbytes)
drv.memcpy_htod(a_gpu, a)
drv.memcpy_htod(b_gpu, b)
# 定义块和网格大小
blocksize = 256
gridsize = (n + blocksize - 1) // blocksize
# 执行核函数
add_vectors = mod.get_function("add_vectors")
add_vectors(
a_gpu, b_gpu, c_gpu, numpy.int32(n), block=(blocksize, 1, 1), grid=(gridsize, 1)
)
# 将结果从 GPU 复制回 CPU
drv.memcpy_dtoh(c, c_gpu)
# 检查计算结果是否正确
assert numpy.allclose(c, a + b), "result not correct"
# 输出结果
print("a:", a)
print("b:", b)
print("c:", c)