How to pass a list of strings to an opencl kernel

问题:

How to pass list of strings to an opencl kernel the right way?

I tried this way using buffers (see following code), but I failed.

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] )

I ran the program first time with "but it failes to output" as 4th list element. Then I played around by increasing and decreasing elements of the list. Finally, there appeared this problem: The output of the program is supposed to be (short version)

(GPU) output : That is VERY cool!

(GPU) output : It is a list!

(GPU) output : A big one!

(GPU) output : WTF is THAT?

But it is:

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. :-(

As you can see, the the 4th list element differes.

So, maybe my approach is wrong or there is a bug in pyopencl or somewhere else.

I am using a NVidia 9400 GPU.

Rambo

回答1:

You code seems to me very complicated. And some part are not very clear to me. For instance, I don't see why you create only one work item:

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

Which force you to loop through your strings (in the kernel) instead of using the available parallelism. Anyhow, if I well understand, you want to send some strings to the GPU copy them in another buffer, get them back in the host side and display them.

If it's the case here is a version using only numpy and of course 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

And the displayed result:

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