When using libjpeg to feed images into OpenCL, to be able to treat channels as normalized uint8's with CL_UNORM_INT8 (floats in the range [0.0, 1.0]), you can only feed it buffers with 4 channel components. This is problematic, because libjpeg only outputs 3 (by default in RGB order) since JPEG has no notion of opacity.
The only workaround I see is to scanlines with libjpeg and then make a duplicate buffer of the appropriate length (with the fourth channel component added for each pixel in the scanlines) and then memcpy the values over, setting the alpha component to 255 for each. You could even do this in place if you are tricky and initialize the buffer to be of row_stride * 4 initially and then walk backwards from index row_stride * 3 - 1 to 0, moving components to the proper places in the full buffer (and adding 255 for alpha where necessary).
However, this feels hacky and if you're dealing with large images (I am), it's unacceptable to have this extra pass over (what will be in aggregate) the entire image.
So, is there a way to get libjpeg to just extend the number of components to 4? I've tried setting properties on cinfo like output_components to no avail. I've read that the only workaround is to compile a special version of libjpeg with the constant RGB_COMPONENTS = 4 set in jmorecfg.h, but this certainly doesn't feel portable or for that matter necessary for such a (common) change of output.
So it turns out that the best solution (at least, the one that doesn't require any custom builds of libs or extra passes through the buffer) is to use libjpeg-turbo. As of 1.1.90 they provide a colorspace constant JCS_EXT_RGBX that adds a fake alpha channel. To my knowledge this is only documented in the release notes of a beta version on SourceForge so barring that this URL changes or no longer exists (read: the internet revolts against sf for its shady insertion of code into "inactive" popular repos and they are forced to shut down), here is the relevant bit reproduced:
When decompressing a JPEG image using an output colorspace of
JCS_EXT_RGBX,JCS_EXT_BGRX,JCS_EXT_XBGR, orJCS_EXT_XRGB, libjpeg-turbo will now set the unused byte to0xFF, which allows applications to interpret that byte as an alpha channel (0xFF= opaque).
Note that this also allows for alternate orderings such as BGR should you need them.
To use it after your jpeg_read_header() call (because this call sets a member on cinfo we need to a default) but before your jpeg_start_decompress() call (because it uses the value of this member), add:
cinfo.out_color_space = JCS_EXT_RGBX; // or JCS_EXT_XRGB, JCS_EXT_BGRX, etc.
And now scanning lines during the decompress will return an extra fourth component for each pixel set to 255.