public Class STBImageResize

Written with emphasis on usability, portability, and efficiency. (No SIMD or threads, so it be easily outperformed by libs that use those.) Only scaling and translation is supported, no rotations or shears. Easy API downsamples w/Mitchell filter, upsamples w/cubic interpolation.

QUICKSTART

stbir_resize_uint8(      input_pixels , in_w , in_h , 0,
                         output_pixels, out_w, out_h, 0, num_channels)
stbir_resize_float(...)
stbir_resize_uint8_srgb( input_pixels , in_w , in_h , 0,
                         output_pixels, out_w, out_h, 0,
                         num_channels , alpha_chan  , 0)
stbir_resize_uint8_srgb_edgemode(
                         input_pixels , in_w , in_h , 0,
                         output_pixels, out_w, out_h, 0,
                         num_channels , alpha_chan  , 0, STBIR_EDGE_CLAMP)
                                                      // WRAP/REFLECT/ZERO

ALPHA CHANNEL

Most of the resizing functions provide the ability to control how the alpha channel of an image is processed. The important things to know about this:

1. The best mathematically-behaved version of alpha to use is called "premultiplied alpha", in which the other color channels have had the alpha value multiplied in. If you use premultiplied alpha, linear filtering (such as image resampling done by this library, or performed in texture units on GPUs) does the "right thing". While premultiplied alpha is standard in the movie CGI industry, it is still uncommon in the videogame/real-time world. If you linearly filter non-premultiplied alpha, strange effects occur. (For example, the average of 1% opaque bright green and 99% opaque black produces 50% transparent dark green when non-premultiplied, whereas premultiplied it produces 50% transparent near-black. The former introduces green energy that doesn't exist in the source image.)
2. Artists should not edit premultiplied-alpha images; artists want non-premultiplied alpha images. Thus, art tools generally output non-premultiplied alpha images.
3. You will get best results in most cases by converting images to premultiplied alpha before processing them mathematically.
4. If you pass the flag FLAG_ALPHA_PREMULTIPLIED, the resizer does not do anything special for the alpha channel; it is resampled identically to other channels. This produces the correct results for premultiplied-alpha images, but produces less-than-ideal results for non-premultiplied-alpha images.
5. If you do not pass the flag FLAG_ALPHA_PREMULTIPLIED, then the resizer weights the contribution of input pixels based on their alpha values, or, equivalently, it multiplies the alpha value into the color channels, resamples, then divides by the resultant alpha value. Input pixels which have alpha=0 do not contribute at all to output pixels unless all of the input pixels affecting that output pixel have alpha=0, in which case the result for that pixel is the same as it would be without FLAG_ALPHA_PREMULTIPLIED. However, this is only true for input images in integer formats. For input images in float format, input pixels with alpha=0 have no effect, and output pixels which have alpha=0 will be 0 in all channels. (For float images, you can manually achieve the same result by adding a tiny epsilon value to the alpha channel of every image, and then subtracting or clamping it at the end.)
6. You can separately control whether the alpha channel is interpreted as linear or affected by the colorspace. By default it is linear; you almost never want to apply the colorspace. (For example, graphics hardware does not apply sRGB conversion to the alpha channel.)