[Pixman] [RFC, PATCH 0/8] Floating point pipeline

[Bill Spitzak]

On 08/25/2012 06:45 PM, Søren Sandmann Pedersen wrote:

> Also note that the linear->sRGB scanline store routine is probably
> slower with these patches because it now uses a binary search instead
> of a simple table lookup. If someone has better suggestions here, I'm
> definitely interested.

I would use a set of bits at the top of the floating point number to 
index a table of linear approximations. Each entry in the table produces 
an A and B floating point value and the result is Ax+B.

Interesting 32-bit floating point numbers in hex:

	0.0	  0x00000000
	0.0031308 0x3b4d2e1c (end of linear part of sRGB)
	1.0       0x3f800000

I would use some if statements to evenly divide the area between just 
before .0031308 and 1.0 into whatever number of segments you want. 
Numbers before this go into a single bin for the linear part of sRGB.
Zero and all negative numbers go into a bin producing a horizontal line 
at zero, and 1.0 and above into a bin producing a horizontal line at 255.

The integer part of the floating point result is the 8-bit number. I 
also very much recommend that the fractional part be used for error 
diffusion, it is added to the next result, and a random starting point 
is chosen for each scan line. This will make gradients look much, much 
nicer.

You can also choose much larger B values such that the resulting 
floating point answers all have the same exponent. This will allow the 
8-bit values to be directly extracted from the resulting floating point.