python - 如何使用 pyopencl 将字符串列表传递给 opencl 内核？

Question

如何以正确的方式将字符串列表传递给 opencl 内核？

我尝试使用缓冲区这种方式(参见下面的代码)，但失败了。

OpenCL(struct.cl):

typedef struct{
          uchar uc[40];
} my_struct9; 

inline void try_this7_now(__global const uchar * IN_DATA ,
                          const uint IN_len_DATA ,
                          __global uchar * OUT_DATA){
    for (unsigned int i=0; i<IN_len_DATA ; i++)  OUT_DATA[i] = IN_DATA[i];
}

__kernel void try_this7(__global const my_struct9 * pS_IN_DATA ,
                        const uint IN_len ,
                        __global my_struct9 * pS_OUT){

    uint idx = get_global_id(0);
for (unsigned int i=0; i<idx; i++) try_this7_now(pS_IN_DATA[i].uc, IN_len, pS_OUT[i].uc);
  }

Python(opencl_struct.py):

# -*- coding: utf-8 -*- 

import pyopencl as cl
import pyopencl.array as cl_array
import numpy

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
# --------------------------------------------------------
LIMIT = 40
mf = cl.mem_flags

import ctypes,sys,struct
"""
typedef struct{
          uchar uc[40];
} my_struct9; 
"""
INlist = []
INlist.append("That is VERY cool!")
INlist.append("It is a list!")
INlist.append("A big one!")
#INlist.append("But it failes to output. :-(")  # PLAY WITH THOSE
INlist.append("WTF is THAT?") # PLAY WITH THOSE
print "INlist : "+str(INlist)
print "largest string "+str( max( len(INlist[iL]) for iL in range(len(INlist)) ) )
strLIMIT=str(LIMIT)
s7 = struct.Struct(  (str(strLIMIT+'s') *len(INlist)) )
IN_host_buffer = ctypes.create_string_buffer(s7.size)
s7.pack_into(IN_host_buffer, 0, *INlist)
IN_dev_buffer = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=IN_host_buffer)

OUT_host_buffer = ctypes.create_string_buffer(s7.size)

OUT_dev_buffer = cl.Buffer(ctx, mf.WRITE_ONLY, len(OUT_host_buffer))
print "> len(OUT_host_buffer) "+str(len(OUT_host_buffer))

# ========================================================================================
f = open("struct.cl", 'r')
fstr = "".join(f.readlines())
prg = cl.Program(ctx, fstr).build()

#cl.enqueue_copy(queue, IN_dev_buffer, IN_host_buffer, is_blocking=True) # copy data to device
cl.enqueue_write_buffer(queue, IN_dev_buffer, IN_host_buffer).wait()

prg.try_this7(queue, (1,), None, IN_dev_buffer, numpy.uint32(LIMIT), OUT_dev_buffer)
# ========================================================================================
cl.enqueue_copy(queue, OUT_host_buffer, OUT_dev_buffer).wait()

SSS = s7.unpack_from(OUT_host_buffer,0)

# unpack here OUT_host_buffer
print "(GPU) output : "+str( SSS )+" "

for s in range(len(SSS)):
 print ">>> (GPU) output : "+str( SSS[s] )

我第一次运行程序时将“但无法输出”作为第四个列表元素。然后我通过增加和减少列表的元素来进行尝试。最后就出现了这个问题: 程序的输出应该是(简短版本)

(GPU) output : That is VERY cool!

(GPU) output : It is a list!

(GPU) output : A big one!

(GPU) output : WTF is THAT?

但事实是:

python opencl_struct.py

INlist : ['That is VERY cool!', 'It is a list!', 'A big one!', 'WTF is THAT?']

largest string 18

len(OUT_host_buffer) 160 (GPU) output : ('That is VERY cool!\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00', 'It is a list!\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00', 'A big one!\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00', 'But it failes to output. :-(\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00')

(GPU) output : That is VERY cool!

(GPU) output : It is a list!

(GPU) output : A big one!

(GPU) output : But it failes to output. :-(

如您所见，第四个列表元素不同。

所以，也许我的方法是错误的，或者 pyopencl 或其他地方有错误。

我使用的是 NVidia 9400 GPU。

兰博

Answer 1

你的代码在我看来非常复杂。有些部分我不是很清楚。例如，我不明白为什么您只创建一个工作项:

prg.try_this7(queue, (1,), None,...)

这迫使您循环遍历字符串(在内核中)，而不是使用可用的并行性。无论如何，如果我很好理解的话，你想发送一些字符串到GPU，将它们复制到另一个缓冲区中，将它们返回到主机端并显示它们。

如果是这种情况，这里是一个仅使用 numpy 的版本，当然还有 pyopencl:

import numpy as np
import pyopencl as cl


ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
#The kernel uses one workitem per char transfert
prog_str = """kernel void foo(global char *in, global char *out, int size){
                  int idx = get_global_id(0);
                  if (idx < size){
                      out[idx] = in[idx];
                  }
           }"""
prog = cl.Program(ctx, prog_str).build()
#Note that the type of the array of strings is '|S40' for the length
#of third element is 40, the shape is 3 and the nbytes is 120 (3 * 40)
original_str = np.array(('this is an average string', 
                         'and another one', 
                         "let's push even more with a third string"))
mf = cl.mem_flags
in_buf = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=original_str)
out_buf = cl.Buffer(ctx, mf.WRITE_ONLY, size=str_size)
copied_str = np.zeros_like(original_str)
#here launch the kernel with str_size number of workitems in this case 120
#this mean that some of the workitems won't process any meaningful char 
#(not all string have a lenght of 40) but it's no biggie
prog.foo(queue, (str_size,), None, in_buf, out_buf, np.int32(str_size))
cl.enqueue_copy(queue, copied_str, out_buf).wait()
print copied_str

显示结果:

['this is an average string' 'and another one'
 "let's push even more with a third string"]