Some simple pyopencl examples
From:
andrew cooke <andrew@...>
Date:
Fri, 30 Sep 2011 17:42:26 -0300
I'm finding pyopencl to be both very cool (so much less code to write) and
very frustrating (such crappy documentation). So here are some completely
basic examples that might help others:
Andrew
def test1():
'''you might think the Array class is something that simplifies code.
this example shows it is not (change the comments and see the errors).
instead, it seems to be something that looks like a numpy array,
but which farms out work to the GPU (operation by operation).'''
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
a = n.array([0], dtype=n.int32)
# this works
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
# this alternative (plus below) fails
#a_dev = cla.to_device(queue, a)
prg = cl.Program(ctx, """
__kernel void test1(__global int* a) {
a[0] = 1;
}
""").build()
event = prg.test1(queue, (1,), None, a_dev)
event.wait()
# this works
cl.enqueue_copy(queue, a, a_dev)
# this alternative(plus above) fails
#a = a_dev.get()
print(a)
def test3():
'''constants (like b) don't need buffering.'''
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
a = n.array([0], dtype=n.int32)
b = n.int32(4)
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
prg = cl.Program(ctx, """
__kernel void test1(__global int* a, const int b) {
a[0] = b;
}
""").build()
event = prg.test1(queue, (1,), None, a_dev, b)
event.wait()
cl.enqueue_copy(queue, a, a_dev)
print(a)
Copying Bytes
From:
andrew cooke <andrew@...>
Date:
Fri, 30 Sep 2011 17:52:05 -0300
def test4():
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
a = n.array([0], dtype=n.int32)
b = n.uint8(129)
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
prg = cl.Program(ctx, """
__kernel void test1(__global int* a, const uchar b) {
a[0] = b;
}
""").build()
event = prg.test1(queue, (1,), None, a_dev, b)
event.wait()
cl.enqueue_copy(queue, a, a_dev)
print(a)
Using Array
From:
andrew cooke <andrew@...>
Date:
Fri, 30 Sep 2011 22:34:11 -0300
With thanks to Bogdan Opanchuk
import pyopencl as cl
import pyopencl.array as cla
import numpy as n
def test5():
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
a = n.array([0], dtype=n.int32)
a_array = cla.to_device(queue, a)
prg = cl.Program(ctx, """
__kernel void test1(__global int* a) {
a[0] = 1;
}
""").build()
event = prg.test1(queue, (1,), None, a_array.data)
event.wait()
a = a_array.get()
print(a)
Struct and packing
From:
andrew cooke <andrew@...>
Date:
Sat, 1 Oct 2011 13:41:06 -0300
import struct as s
import pyopencl as cl
import numpy as n
def test6():
# i have intel and amd installed (running on cpu).
# switching gives different error messages (useful at times!)
# p = cl.get_platforms();
# print(p)
# d = p[0].get_devices() # 1 is amd
# print(d)
# ctx = cl.Context(devices=d)
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
for use_struct in (True, False):
if use_struct:
a = s.pack('=ii',1,2)
print(a, len(a))
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, len(a))
else:
a = n.array([(1,2)], dtype=n.dtype('2i4', align=True))
# a = n.array([(1,2)], dtype=n.dtype('2i4'))
print(a, a.itemsize, a.nbytes)
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
b = n.array([0], dtype='i4')
print(b, b.itemsize, b.nbytes)
b_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, b.nbytes)
c = n.array([0], dtype='i4')
print(c, c.itemsize, c.nbytes)
c_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, c.nbytes)
prg = cl.Program(ctx, """
typedef struct s {
int f0;
int f1 __attribute__ ((packed));
} s;
__kernel void test(__global const s *a, __global int *b, __global
int *c) {
*b = a->f0;
*c = a->f1;
}
""").build()
cl.enqueue_copy(queue, a_dev, a)
event = prg.test(queue, (1,), None, a_dev, b_dev, c_dev)
event.wait()
cl.enqueue_copy(queue, b, b_dev)
print(b)
cl.enqueue_copy(queue, c, c_dev)
print(c)
if __name__ == '__main__':
test6()
Bytes in struct
From:
andrew cooke <andrew@...>
Date:
Sat, 1 Oct 2011 13:45:06 -0300
Like taking candy from a baby :o) (note that writing to bytes is an oencl
extension, which is why I am using ints as output).
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
for use_struct in (True, False):
if use_struct:
a = s.pack('=bb',1,2)
print(a, len(a))
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, len(a))
else:
# a = n.array([(1,2)], dtype=n.dtype('2i1', align=True))
a = n.array([(1,2)], dtype=n.dtype('2i1'))
print(a, a.itemsize, a.nbytes)
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
b = n.array([0], dtype='i4')
print(b, b.itemsize, b.nbytes)
b_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, b.nbytes)
c = n.array([0], dtype='i4')
print(c, c.itemsize, c.nbytes)
c_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, c.nbytes)
prg = cl.Program(ctx, """
typedef struct s {
char f0;
char f1 __attribute__ ((packed));
} s;
__kernel void test(__global const s *a, __global int *b, __global
int *c) {
*b = a->f0;
*c = a->f1;
}
""").build()
cl.enqueue_copy(queue, a_dev, a)
event = prg.test(queue, (1,), None, a_dev, b_dev, c_dev)
event.wait()
cl.enqueue_copy(queue, b, b_dev)
print(b)
cl.enqueue_copy(queue, c, c_dev)
print(c)
Trickier Alignment
From:
andrew cooke <andrew@...>
Date:
Sat, 1 Oct 2011 13:50:40 -0300
This shows the importance of numpy's "align" keyword.
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
for use_struct in (True, False):
if use_struct:
a = s.pack('=bi',1,2)
print(a, len(a))
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, len(a))
else:
a = n.array([(1,2)], dtype=n.dtype('i1i4', align=True))
# this no longer works - without align=True we get the wrong value
# a = n.array([(1,2)], dtype=n.dtype('i1i4'))
print(a, a.itemsize, a.nbytes)
a_dev = cl.Buffer(ctx, cl.mem_flags.WRITE_ONLY, a.nbytes)
b = n.array([0], dtype='i4')
print(b, b.itemsize, b.nbytes)
b_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, b.nbytes)
c = n.array([0], dtype='i4')
print(c, c.itemsize, c.nbytes)
c_dev = cl.Buffer(ctx, cl.mem_flags.READ_ONLY, c.nbytes)
prg = cl.Program(ctx, """
typedef struct s {
char f0;
int f1 __attribute__ ((packed));
} s;
__kernel void test(__global const s *a, __global int *b, __global
int *c) {
*b = a->f0;
*c = a->f1;
}
""").build()
cl.enqueue_copy(queue, a_dev, a)
event = prg.test(queue, (1,), None, a_dev, b_dev, c_dev)
event.wait()
cl.enqueue_copy(queue, b, b_dev)
print(b)
cl.enqueue_copy(queue, c, c_dev)
print(c)
thanks a bunch
From:
Ajay Shah <ajayshah.mcmaster@...>
Date:
Mon, 3 Sep 2012 12:42:48 -0400
Hey andrew,
Just wanted to say thanks for the examples. I've been trying to figure it
out, but with little luck. I'll take a better look at your examples later
and hope that things workout.
btw, what operating system did you use to compile the code?
i was wondering, if I had some questions in the near future, could I run
them past you?
I hope that you will still post more examples,
thanks again,
Ajay
Crappy docs
From:
Andreas Kloeckner <lists@...>
Date:
Tue, 16 Oct 2012 03:31:04 -0400
Hi Andrew,
Andreas here, author of PyOpenCL. I'd love to hear what you've found
frustrating about the PyOpenCL docs. While I've got limited time to
spend on the project, I'd still love to try and make this better.
Andreas
Re: Crappy docs
From:
andrew cooke <andrew@...>
Date:
Tue, 16 Oct 2012 08:14:12 -0300
I don't remember PyOpenCL specific details, but from the examples above (and
my thoughts below on PyCUDA) I imagine that it's a lack of simple,
self-contained examples that show the basic functionality of the different
components.
The aim is not to teach how to use OpenCL, but how to do basic things in
OpenCL with that library.
These could be tests (more strongly - should be run with tests, because
otherwise there's no guarantee they work after a new release) BUT with tests
it's common that you end up with a very abstract framework that avoid
repetition and allows you to test all variations. In contrast these examples
would repeat a lot of basic book-keeping, just so that each is self-contained.
I've seen some projects use a wiki for this kind of thing, with people
contributing, but I am not sure your project is large enough to support that.
More recently I have been using PyCUDA (different job / client) and the same
is true there. Some examples of the things I would like (which are perhaps a
little more advanced) include:
- Showing how different kernels can be sscheduled to run after each other
or in parallel.
- Showing how to mix low level work (hand-written kernels) with high level
work (eg calling one of the FFT libraries).
- A list of recommended libraries (eg for FFT).
Again, the format would be complete, self-contained examples. As far as I
remember, every object / method has some docs, but the higher level picture is
missing.
Cheers,
Andrew
Re: Crappy docs
From:
andrew cooke <andrew@...>
Date:
Tue, 16 Oct 2012 17:01:51 -0300
Like the wiki, yes. I am surprised / worried I didn't find those when working
with PyCUDA. Perhaps I assumed I wouldn't after using PyOpenCL?
The other ones look too much like tests (see my prev comment). I realise that
you could pull them apart and find useful info, but if you're just poking
around looking for something that "looks like an example" I think you'd could
easily miss hwat is there. Compare the array one with what I wrote, which
includes a human-readable comment that gives some context about what an array
is meant to do.
Cheers,
Andrew
Re: Crappy docs
From:
Andreas Kloeckner <lists@...>
Date:
Sat, 20 Oct 2012 00:09:42 -0400
andrew cooke <andrew@...> writes:
> Like the wiki, yes. I am surprised / worried I didn't find those when working
> with PyCUDA. Perhaps I assumed I wouldn't after using PyOpenCL?
>
> The other ones look too much like tests (see my prev comment). I realise that
> you could pull them apart and find useful info, but if you're just poking
> around looking for something that "looks like an example" I think you'd could
> easily miss hwat is there. Compare the array one with what I wrote, which
> includes a human-readable comment that gives some context about what an array
> is meant to do.
Thanks for this feedback. I'll try to keep it in mind going forward.
Andreas
Comment on this post