Hi all, I've been playing with integrating Tesseract with Ghostscript for the past couple of weeks. I have it working nicely, and I've started trying to pass back some of the tweaks I've done along the way as pull requests on github - thanks to stweil for bearing with me!
The biggest of these is an implementation of matrixDotVector for NEON equipped ARMs (intsimdmatrixneon.cpp). This makes a massive difference to the speed on ARM devices (such as my Raspberry pi), but (depending on what language data I feed it), profiles still show 30-55% of runtime still in this function. It'd be nice to do the whole thing in NEON, but NEON doesn't support doubles, so we have to drop back to "standard" operations to add the biases/apply the scales. If we were using floats, we'd be golden. It's possible that the calling code could be tweaked to use floats instead of doubles. So, before I dive into this, I thought I'd ask here. Presumably, there is a good reason why the existing code uses doubles rather than floats? Am I doomed to damage the quality of the results I get out by moving to floats? Thanks in advance for any help/insight people can offer. -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-ocr+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/0e4552a1-a0d3-4a38-8881-0972bdf3185b%40googlegroups.com.
