Question

我想在大型2D复杂阵列上运行一个函数（最终2 * 12x2 * 12数据点）。但是，pycuda不能按预期工作。 ElementWise函数在2d数组中不起作用，因此我使用了具有块大小的SourceModule函数。

现在的问题是GPU上的C代码与CPU上的numpy计算结果不一致。结果非常大而奇怪。

我正在使用以下代码。出了什么问题？

#!/usr/bin/env python
#https://github.com/lebedov/scikits.cuda/blob/master/demos/indexing_2d_demo.py
"""
Demonstrates how to access 2D arrays within a PyCUDA kernel in a
numpy-consistent manner.
"""

from string import Template
import pycuda
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
from pycuda.compiler import SourceModule
import numpy as np
from matplotlib import pyplot as plt

# Set size
A = 2**3
B = 2**3
N = A*B
x_gpu = gpuarray.to_gpu(np.fromfunction(lambda x,y: (1.+x)*np.exp(1.j*y*np.pi/10), (A,B)) )
y_gpu = gpuarray.to_gpu(np.fromfunction(lambda x,y: 1.*x, (A,B)).astype(
                                x_gpu.dtype)) 
d_gpu = gpuarray.to_gpu(np.zeros_like(x_gpu.get()))#.astype(np.float32))

func_mod_template = Template("""
// Macro for converting subscripts to linear index:
#define INDEX(a, b) a*${B}+b
#include <pycuda-complex.hpp>

//__global__ void func(double *d,double *x,double *y, unsigned int N) {
__global__ void func(pycuda::complex<float> *d,pycuda::complex<float> *x,
                     pycuda::complex<float> *y)
{
    // Obtain the linear index corresponding to the current thread:     
    // unsigned int idx =  blockIdx.y*blockDim.y*gridDim.x + 
                        blockIdx.x*blockDim.x*gridDim.y +threadIdx.x+threadIdx.y;
    unsigned int block_num        = blockIdx.x + blockIdx.y * gridDim.x;              
    unsigned int thread_num       = threadIdx.y * blockDim.x + threadIdx.x;           
    unsigned int threads_in_block = blockDim.x * blockDim.y;                          
    unsigned int idx              =  (threads_in_block * block_num + thread_num);

    // Convert the linear index to subscripts:
    unsigned int a = idx/${B};
    unsigned int b = idx%${B};

    // Use the subscripts to access the array:
    // d[INDEX(a,b)] = x[INDEX(a,b)]+y[INDEX(a,b)];
    pycuda::complex<float> j(0,arg(x[idx]));
    pycuda::complex<float> i(abs(x[idx]),0);
    d[idx] = i * exp(j);
}
""")

max_threads_per_block = pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_THREADS_PER_BLOCK)
max_block_dim = (pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_BLOCK_DIM_X),
                 pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_BLOCK_DIM_Y),
                 pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_BLOCK_DIM_Z))
max_grid_dim = (pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_GRID_DIM_X),
                pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_GRID_DIM_Y),
                pycuda.autoinit.device.get_attribute(pycuda._driver.device_attribute.MAX_GRID_DIM_Z))
max_blocks_per_grid = max(max_grid_dim)
block_dim = max_block_dim
block_dim = (max_block_dim[0],1,1)
grid_dim = (int(np.ceil(1.*x_gpu.shape[0]/block_dim[0])),
            int(np.ceil(1.*x_gpu.shape[1]/block_dim[1])))
print block_dim,grid_dim, N

func_mod = \
         SourceModule(func_mod_template.substitute(max_threads_per_block=max_threads_per_block,
                                                   max_blocks_per_grid=max_blocks_per_grid,
                                                   A=A, B=B))
func = func_mod.get_function('func')
func(d_gpu,x_gpu,y_gpu,
     block=block_dim,
    grid=grid_dim)

print d_gpu.get()/x_gpu.get()
#print 'Success status: ', np.allclose(x_np, x_gpu.get())
plt.imshow((d_gpu.get()/x_gpu.get()).real)
plt.colorbar()
plt.show()

Answer 1

作为实际答案：将x_gpu行更改为

x_gpu = gpuarray.to_gpu(np.fromfunction(
    lambda x,y: (1.+x)*np.exp(1.j*y*np.pi/10), (A,B)).astype(np.complex64) )

似乎解决了这个问题。此外，虽然ElementwiseKernel不适用于2D阵列，但无论如何你都在使用2d-> 1d转换，所以没有什么能阻止你写作

func = ElementwiseKernel(
    "pycuda::complex<float> *d, pycuda::complex<float> *x, pycuda::complex<float> *y",

    Template("""
    // Convert the linear index to subscripts:
    unsigned int a = i/${B};
    unsigned int b = i%${B};

    // Use the subscripts to access the array:
    //d[INDEX(a,b)] = x[INDEX(a,b)]+y[INDEX(a,b)];
    pycuda::complex<float> angle(0,arg(x[i]));
    pycuda::complex<float> module(abs(x[i]),0);
    d[i] = module * exp(angle);
    """).substitute(A=A, B=B),

    preamble=Template("""
    #define INDEX(a, b) a*${B}+b
    """).substitute(A=A, B=B))

func(d_gpu, x_gpu, y_gpu)

这样您就不需要处理块/网格大小，因为PyCUDA会为您处理此问题。

pycuda :: complex数组的元素智能函数

1 个答案: