Question

In OpenCL you need to define a sampler objected when doing image reads (read_image*), and in that sampler you specify how addresses are clamped (e.g. you can do CLK_ADDRESS_CLAMP_TO_EDGE). But there is no sampler for image writes (write_image*). So there is no clamping specified.
What is supposed to happen to pixel addresses that exit the image boundry then? Is that just not specified and left up to the vendor?
Looking on the web, it seems most OpenCL code that has write_images either does not use any clamping or people use their own clamping functions. I don’t understand why the clamping is unspecified for writes, but specified for reads. Am I missing something here? I’ve googled and double read the specs, there’s nothing explaining how writes are clamped.

Answer 1

From the OpenCL 1.2 specs, 6.12.14.4:

The behavior of write_imagef, write_imagei and write_imageui for image
objects created with image_channel_data_type values not specified in
the description above or with (x, y) coordinate values that are not in
the range (0 … image width – 1, 0 … image height – 1),
respectively, is undefined.

There is no write clamping, and it’s left to the user kernel code to ensure coordinates are valid.