Add some documentation.

docs/bilinear_interpolation.md

@@ -0,0 +1,59 @@
+## Coding Bilinear Interpolation ##
+
+Source: [The Supercomputing Blog](http://supercomputingblog.com/graphics/coding-bilinear-interpolation/)
+
+When performing image transformation and manipulation techniques, it is often
+necessary to employ some sort of interpolation or filtering in order to obtain
+a good image quality. For example, if you scale an image, you can determine
+the final color of each pixel by either some basic nearest neighbor method, or
+a more advanced interpolation method. However, an interpolation method will
+invariably offer better final image quality. In this tutorial, we’ll be
+writing a function to rotate an image, using bilinear interpolation. This
+tutorial also demonstrates how to perform a high quality image rotate
+transformation, however, that is not the focus of this tutorial, but rather
+the example transform being performed.
+
+
+### Why interpolation is used after image transforms ###
+
+Image transforms like scaling, rotating, twisting or otherwise warping,
+effectively move pixels around in the image. The result is that the pixels in
+the final image do not directly map to a single original pixel in the original
+image. Suppose after rotation, a pixel should have the same color as pixel
+(3.4, 5.6) in the original image. The only problem is that there is no such
+thing as fractional pixels! Instead of simply choosing the color of the
+nearest available pixel, (3,6), we can get better results by intelligently
+averaging the colors of the four closest pixels. In this case, the values of
+pixels (3,5) (3,6) (4,5) and (4,6) would be intelligently averaged together to
+estimate the color of an imaginary pixel at (3.4, 5.6). This is called
+interpolation, and it is used after image transforms to provide a smooth,
+accurate and visually appealing images.
+
+
+### Understanding bilinear interpolation ###
+
+We can best understand bilinear interpolation by looking at the graphic here.
+The green P dot represents the point where we want to estimate the color. The
+four red Q dots represent the nearest pixels from the original image. The
+color of these four Q pixels is known. In this example, P lies closest to Q12,
+so it is only appropriate that the color of Q12 contributes more to the final
+color of P than the 3 other Q pixels.
+
+
+### Calculating bilinear interpolation ###
+
+There are several ways equivalent ways to calculate the value of P. An easy
+way to calculate the value of P would be to first calculate the value of the
+two blue dots, R2, and R1. R2 is effectively a weighted average of Q12 and
+Q22, while R1 is a weighted average of Q11 and Q21.
+
+``` R1 = ((x2 – x)/(x2 – x1))*Q11 + ((x – x1)/(x2 – x1))*Q21 R2 = ((x2 –
+x)/(x2 – x1))*Q12 + ((x – x1)/(x2 – x1))*Q22 ```
+
+After the two R values are calculated, the value of P can finally be
+calculated by a weighted average of R1 and R2.
+
+``` P = ((y2 – y)/(y2 – y1))*R1 + ((y – y1)/(y2 – y1))*R2 ```
+
+The calculation will have to be repeated for the red, green, blue, and
+optionally the alpha component of.

docs/padding.txt

@@ -0,0 +1,16 @@
+https://stackoverflow.com/questions/11642210/computing-padding-required-for-n-byte-alignment
+
+There is a faster way to compute the padding, if the alignment is a power of two
+(2,4,8,...). The following runs because binary & is similar to % for powers of
+two: %(2^x) and &(2^x-1) do the same for positive numbers. Attention: & will
+delete the sign bit and therefore always returns the positive modulo result.
+
+So (4 - (string_length & 3)) & 3 will do the same as
+(4 - (string_length % 4)) % 4. Using the positive modulo property this can be
+simplified to (-string_length) & 3!
+
+
+If you wanna add that result to the size you can even do more optimizations:
+
+padded_length = (string_length + 3) & ~3 Semantically this 'rounds up' the
+number to the padding size of 4.

docs/srgb_linear.md

@@ -0,0 +1,72 @@
+## A close look at the sRGB formula ##
+
+[Source](https://blog.demofox.org/2018/03/10/dont-convert-srgb-u8-to-linear-u8/)
+[Specific page](http://entropymine.com/imageworsener/srgbformula/)
+
+The formulas for converting sRGB colors to linear, and the reverse, are usually
+given as follows:
+
+```
+sRGB to Linear
+
+Condition            Value
+0 ≤ S ≤ 0.04045      L = S/12.92
+0.04045 < S ≤ 1      L = ((S+0.055)/1.055)^2.4
+```
+
+```
+Linear to sRGB
+
+Condition            Value
+0 ≤ L ≤ 0.0031308    S = L×12.92
+0.0031308 < L ≤ 1    S = 1.055×L^(1/2.4) − 0.055
+```
+
+The definition uses two different functions – a straight line and an
+exponential curve – glued together at a certain “cutoff point”. The implication
+is that these functions (the ones in the sRGB to Linear definition) intersect
+at the point:
+
+**(0.0404500000000000, 0.00313080000000000)**
+
+If we take a close look at the function, though, we can see this intersection
+point is not exact, so the formula has a slight discontinuity:
+
+* sRGB: 0.0404500000000000
+* sRGB: 0.0404500000000001
+* Linear: 0.00313080495356037
+* Linera: 0.00313080728306769
+
+If we extend the functions, we see that they actually intersect at two points:
+
+1. (0.0381547987133173, 0.00295315779514840)
+2. (0.0404482362771082, 0.00313066844250063)
+
+If we use either of these two points as the cutoff point, there will be no
+discontinuity at all. It’s apparent that the sRGB inventors selected the second
+of these points, and rounded its value to 4 significant digits to get 0.04045.
+They then used that rounded sRGB value to recalculate the Linear value, and got
+0.00313080. I don’t know why their Linear value has more significant digits (5
+or 6) than their sRGB value (4).
+
+I guess that if you’re a perfectionist, and want no discontinuity at all, you
+can tweak the formulas like this:
+
+```
+sRGB to Linear
+
+Condition                       Value
+0 ≤ S ≤ 0.0404482362771082      L = S/12.92
+0.0404482362771082 < S ≤ 1      L = ((S+0.055)/1.055)^2.4
+```
+
+```
+Linear to sRGB
+
+Condition                       Value
+0 ≤ L ≤ 0.00313066844250063     S = L×12.92
+0.00313066844250063 < L ≤ 1     S = 1.055×L^(1/2.4) − 0.055
+```
+
+It doesn’t make any practical difference. The two functions are extremely close
+together in this region, so the precise cutoff point is almost irrelevant.