pyopencl global_work_offset 内核参数

pyopencl global_work_offset kernel argument

我想使用来自 OpenCL API 函数 clEnqueueNDRangeKernel. I can't figure out how to do that within pyopencl API 的 global_work_offset 参数。这是一个演示代码,我想在其中向内核调用添加偏移量 2,因此 get_global_id(0) 从 2 而不是 0 开始:

import pyopencl as cl 
import pyopencl.array 
import numpy as np

platform = cl.get_platforms()[0]
devices = platform.get_devices()[1] #gpu
context = cl.Context(devices=[devices])
queue =  cl.CommandQueue(context)

kernel = cl.Program(context, """
    __kernel void derp(global char* a) {
        a[get_global_id(0)] = 1;
    }""").build()

buffarr = cl.array.zeros(queue, 4, dtype=np.uint8)
kernel.derp(queue, (2,), None, buffarr.data)

np_data = buffarr.get()

# within this demo the buffer contains currently [1,1,0,0]
assert np.array_equal(np_data, [0,0,1,1])

如何更改代码以使断言不会失败?我不想在此处向内核代码添加额外的参数。

作为 documentation,您可以将 global_offset 作为命名参数传递。

内核的调用变为:

kernel.derp(queue, (4, 1), None, buffarr.data, global_offset=[2, 0])

更改的程序:

import pyopencl as cl
import pyopencl.array
import numpy as np


platform = cl.get_platforms()[2]
print(platform)
devices = platform.get_devices()[0] #gpu
context = cl.Context(devices=[devices])
queue =  cl.CommandQueue(context)

kernel = cl.Program(context, """
    __kernel void derp(global char* a) {
        a[get_global_id(0)] = 1;
    }""").build()


buffarr = cl.array.zeros(queue, 4, dtype=np.uint8)

# (4, 1) ==> shape of the buffer
kernel.derp(queue, (4, 1), None, buffarr.data, global_offset=[2, 0])

np_data = buffarr.get()
print(np_data)
# within this demo the buffer contains currently [1,1,0,0]
assert np.array_equal(np_data, [0,0,1,1])
print("Ok")

执行后:

在设备 0 上

<pyopencl.Platform 'Intel(R) OpenCL' at 0x60bdc0>
[0 0 1 1]
Ok

在设备 1 上

<pyopencl.Platform 'Experimental OpenCL 2.0 CPU Only Platform' at 0xb60a20>
[0 0 1 1]
Ok

在设备 2 上

<pyopencl.Platform 'NVIDIA CUDA' at 0xff0440>
[0 0 1 1]
Ok

测试 python 2.7.11 [MSC v.1500 64 位 (AMD64)] - pyopencl (2015, 1)