Performance Tuning Tips

IVL considers the lower-left corner of an image to be the origin. Images are rendered from the bottom row, up. To avoid the need for data conversion by the application program, store images in memory so the pixel at the lower-left corner is the first pixel in memory, followed by the remainder of the bottom row, then subsequent rows from the bottom to the top of the image.

Angles of rotation will round to the nearest 1/10th degree. This is true for both the IVX and the software rendering paths.

Do not arbitrarily intermix IVL and X procedure calls.

For flicker-free, near-real-time, display of annotation, use the overlay visual and only change text that requires changing.

Do not use glXMakeCurrent needlessly for every image frame. In other words, don't call glXMakeCurrent unless you are changing contexts.

IVX supports X and Y scale values with absolute values between 0.75 and 32.0.

IVX supports image translation values within the following limits:

Image sizes should be at least 8×4 pixels (width×height), and at most 32768×32768 pixels.

Only three clipping rectangles (including any enabled scissor specification) are supported in hardware. Additional clipping rectangles will cause rendering to be accomplished with multiple rendering passes through the hardware, so rendering performance may decease by more than half.

Nearest neighbor interpolation operates faster than bilinear interpolation, which in turn operates faster than bicubic interpolation. So, a progressive refinement scheme will provide the best results where interactivity is needed.

Rotations that alter the x-axis direction will have slower performance than orientations that maintain a positive sense for x (such as no rotation, or a flip about the x axis using a negative y scalar). If your data is typically oriented to require rotating, consider reorienting the data instead of using image transformation to perform this operation.

A scaling factor of 2.0 for both x and y (that is, 2× zoom) with bilinear resampling for luminance data performs better than any other image transformation.

Avoid clearing a single buffer, and clearing and swapping double-buffers if not necessary.