2009/2/13 Yang Zhao :

2009/2/13 Matthias Hopf :

...

On Feb 11, 09 11:28:56 -0800, Yang Zhao wrote:

...

static float ps_alu_consts[] = { - 1.0, 0.0, 1.13983, -1.13983/2, // r - c[0] - 1.0, -0.39465, -0.5806, (0.39465+0.5806)/2, // g - c[1] - 1.0, 2.03211, 0.0, -2.03211/2, // b - c[2] + 1.0, 0.0, 1.4020, 0, // r - c[0] + 1.0, -0.34414, -0.71414, 0, // g - c[1] + 1.0, 1.7720, 0.0, 0, // b - c[2] + /* Constants for undoing Y'CbCr scaling + * - Y' is scaled from 16:235 + * - Cb/Cr are scaled from 16:240 + * Unscaled value N' = N * N_mul + N_shift (N' in range [-0.5, 0.5]) + * Vector is [Y_mul, Y_shfit, C_mul, C_shift] + */ + 256.0/219.0, -16.0/219.0, 256.0/224.0, -128.0/224.0,

I don't understand this, Alex. You are changing *both* the color conversion matrix *and* introducing scaling functions?!?

The change to the conversion matrix is actually to make it follow BT.709 instead of BT.601. The former seemed like a better default, as it's what's used in HD media, and the difference between the two is claimed to be mostly negligible.

Ops, that's not quite true either. What was there before was Y'UV to RGB. What we actually want for digital media is Y'CbCr to RGB, which is ambiguously (and somewhat incorrectly) labelled as "YUV" in Xv. The new version is with coefficients for BT.601. -- Yang Zhao http://yangman.ca -- To unsubscribe, e-mail: radeonhd+unsubscribe@opensuse.org For additional commands, e-mail: radeonhd+help@opensuse.org

On Feb 13, 09 09:29:45 -0800, Yang Zhao wrote:

2009/2/13 Matthias Hopf :

...

I don't understand this, Alex. You are changing *both* the color conversion matrix *and* introducing scaling functions?!?

The change to the conversion matrix is actually to make it follow BT.709 instead of BT.601. The former seemed like a better default, as

Ah, I see. Then everything makes sense. Alex, Yang, sorry that I apparently mixed you both up.

it's what's used in HD media, and the difference between the two is claimed to be mostly negligible.

Which is interesting, because the coefficients are actually pretty different.

...

The scaling can be perfectly embedded in the color matrix.

My linear algebra is a little rusty, so I wasn't sure if was possible to fold the 2 sets scaling multiplier and offset into the conversion matrix when I made the patch. I can give this another look-over.

;-) It's possible with 4x4 matrices, and input vectors (read: color) with a 1.0 in the forth component. Analogous to homogeneous coordinates. Matthias -- Matthias Hopf __ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de -- To unsubscribe, e-mail: radeonhd+unsubscribe@opensuse.org For additional commands, e-mail: radeonhd+help@opensuse.org