Search code examples
pythonimagenumpyopenclpyopencl

PyopenCL 3D RGBA image from numpy array


I want to construct an OpenCL 3D RGBA image from a numpy array, using pyopencl. I know about the cl.image_from_array() function, that basically does exactly that, but doesn't give any control about command queues or events, that is exposed by cl.enqueue_copy(). So I really would like to use the latter function, to transfer a 3D RGBA image from host to device, but I seem to not being able getting the syntax of the image constructor right.

So in this environment

import pyopencl as cl
import numpy as np

platform = cl.get_platforms()[0]
devs = platform.get_devices()
device1 = devs[1]
mf = cl.mem_flags
ctx = cl.Context([device1])
Queue1=cl.CommandQueue(ctx,properties=cl.command_queue_properties.PROFILING_ENABLE)

I would like to do something analog to

  d_colortest = cl.image_from_array(ctx,np.zeros((256,256,256,4)).astype(np.float32),num_channels=4,mode='w')

Using the functions

d_image = cl.Image(...)
event = cl.enqueue_copy(...)

Solution

  • I adapted the cl.image_from_array() function to be able to return an event, which was basically straightforward:

    def p_Array(queue_s, name, ary, num_channels=4, mode="w", norm_int=False,copy=True):
        q = eval(queue_s)
        if not ary.flags.c_contiguous:
            raise ValueError("array must be C-contiguous")
    
        dtype = ary.dtype
        if num_channels is None:
    
            from pyopencl.array import vec
            try:
                dtype, num_channels = vec.type_to_scalar_and_count[dtype]
            except KeyError:
                # It must be a scalar type then.
                num_channels = 1
    
            shape = ary.shape
            strides = ary.strides
    
        elif num_channels == 1:
            shape = ary.shape
            strides = ary.strides
        else:
            if ary.shape[-1] != num_channels:
                raise RuntimeError("last dimension must be equal to number of channels")
    
            shape = ary.shape[:-1]
            strides = ary.strides[:-1]
    
        if mode == "r":
            mode_flags = cl.mem_flags.READ_ONLY
        elif mode == "w":
            mode_flags = cl.mem_flags.WRITE_ONLY
        else:
            raise ValueError("invalid value '%s' for 'mode'" % mode)
    
        img_format = {
                1: cl.channel_order.R,
                2: cl.channel_order.RG,
                3: cl.channel_order.RGB,
                4: cl.channel_order.RGBA,
                }[num_channels]
    
        assert ary.strides[-1] == ary.dtype.itemsize
    
        if norm_int:
            channel_type = cl.DTYPE_TO_CHANNEL_TYPE_NORM[dtype]
        else:
            channel_type = cl.DTYPE_TO_CHANNEL_TYPE[dtype]
    
        d_image = cl.Image(ctx, mode_flags,
                cl.ImageFormat(img_format, channel_type),
                shape=shape[::-1])
        if copy:
            event = cl.enqueue_copy(q,d_image,ary,origin=(0,0,0),region=shape[::-1])
            event_list.append((event,queue_s,name))
        return d_image, event