More keyword changes: sobelN_a etc.,

timhutton commented 3 years ago

New stencils: sobelN_a, sobelNE_a etc. (all 8 for completeness)
Formulas now support a_e, a_nw etc. in a more natural way. Instead of accessing the float4 block that is the neighbor of the current block, these keywords return a float4 that has the four values that are neighbors of the cells. Now we can do e.g. a = a_ne; in a formula to have the whole pattern shift down and left by one pixel. This means there is much less swizzling to do to collect the cells you want in a float4 - we've done the swizzling for you.
Updated patterns to use the new features - see the files changed.
Added documentation in Help/writing_new_rules.html to explain it.

timhutton commented 3 years ago

@danwills Good point about smoke.vti, it looks like it will be fine as a formula. I'll give it a go.

If there's a way for wave_equation.vti to be a single reagent that would be great but I don't think there is - the rate of change of a is an independent quantity and thus must be stored somewhere. Consider the case where a is all zero and b has an initial pattern.

(Of course you could interleave the a and b cells, or do similar tricks, but we don't want that.)

I guess we can get rid of the time delay in wave_equation.vti by updating b first:

b += laplacian_a * timestep;
delta_a = b;

And in fact this seems to be a better way of doing things - it now tolerates the damping being zero and the timestep can be larger. I'll update it!

timhutton commented 3 years ago

I've updated smoke.vti to be a formula and use the a_n etc. keywords. Let's look at the stencils it could use.

v combines input from a and b:

float4 v  = a_n - a_s - b_e + b_w + sq2 * (a_nw + b_nw) + sq2 * (a_ne - b_ne) + sq2 * (b_sw - a_sw) - sq2 * (b_se + a_se);

You could separate this into a and b inputs:

float4 v  = a_n - a_s + sq2 * (a_nw + a_ne - a_sw - a_se);
v  += - b_e + b_w + sq2 * (b_nw - b_ne + b_sw - b_se);

which would then be stencils like this:

 sq2     1    sq2             sq2     0   -sq2
  0      0     0   * a         1      0    -1   * b
-sq2    -1   -sq2             sq2     0   -sq2

but you'd have to separate out the sq2 values to avoid needing to hard-code them into the stencil.

So if we had stencils like this:

                      1  0  1
y_corner_gradient =   0  0  0
                     -1  0 -1

and

                      1  0 -1
x_corner_gradient =   0  0  0
                      1  0 -1

we could write something like:

float4 v = sq2 * y_corner_gradient_a + y_gradient_a + sq2 * x_corner_gradient_b - x_gradient_b;

I dont want to do this immediately but it certainly could be done for frequent use-cases.