Yo there.

I'm facing weird problem with my normals, so if someone finds a solution please share!

Desc:
I'm testing two identical grid-meshes, based on heightmap 256x256.
One is computed at runtime in vertex shader, where height is taken through vft.
The second one is generated on cpu with heights taken directly from texture data.
Normals are generated with d3dxcomputenormals, so the second mesh is lit properly by simple diffuse directional light.
For testing purposes I do output dx normals as colour, so I see nice green/red/magenta colours:)

Next... All the normals from the cpu mesh were saved in texture (A8R8G8B8) in the form of:
(vertexnormal * 0.5 + 0.5) * 255;
Then, on the other side, deep in the VS I take them out with tex2D(norms,proper_coord) * 2 - 1;

I checked encoding routine many times, assuring that proper channels are filled and still I can't get proper normals. They are
almost the same, but I see the lack of blue/magenta colours.

What I want to do is to just check the method of passing normals through texture and heck I don't see that invisible misstake or error:P
From what I wrote, do you see things that might be destroying my results?

When I'll be at home in six hours I'll make some additional shots and attach them. Regarding the byte order, I think I checked that already, in some cases swaping R with G was helping to obtain similar results, but still lacking blue:( If I won't fix it, I'll render to texture where I'll be certain that I pass proper vals:P

EDIT: Ok, changed few things and situation is a bit different.
There's a screenshot showing the state of normalmap and vertex normals.
Here's some data got from mesh vertex buffer (in most cases vertex normals are close to 1,0,0):

VERTEXNORMAL: 0.92, 0.29, 0.00
ENCODED: 0.96, 0.64, 0.50 (normal * 0.5 + 0.5)
DECODED: 0.92, 0.29, 0.00 (fetchnormal * 2 - 1)
TEXCOLOR: 244.76, 164.29, 127.50 (it goes to external texture file)

VERTEXNORMAL: 0.55, 0.18, 0.00
ENCODED: 0.78, 0.59, 0.50
DECODED: 0.55, 0.18, 0.00
TEXCOLOR: 198.24, 151.06, 127.50
...
I don't really know in what space d3dxcompute calculates normals so they should point 0,1,0 instead of 1,0,0:P

Bueh!

Subject might be closed or deleted, useless problem...
But before, could someone explain me why in FVF format NORMAL has to be before DIFFUSE?
I found out that normal is saved in bgra format, but I was still missing one component, the Z, always equal to zero.
So I remember I had problems with fvf before I tried to replace fvf and vertex data and suddenly the lost val has appeared.
I can't believe it:P