NOTE: GLK's approximate ranking of 8 most important tagged with
[GLK:1], [GLK:2], ...
========================
SHORT TERM ============= (*needed* for streamlines & tractography)
========================
Remove CL from compiler
[GLK:3] Add sequence types (needed for evals & evecs)
syntax
types: ty '{' INT '}'
value construction: '{' e1 ',' … ',' en '}'
indexing: e '{' e '}'
[GLK:4] evals & evecs for symmetric tensor[2,2] and
tensor[3,3] (requires sequences)
ability to emit/track/record variables into dynamically re-sized
runtime buffer
tensor fields: convolution on general tensor images
========================
SHORT-ISH TERM ========= (to make using Diderot less annoying to
======================== program in, and slow to execute)
value-numbering optimization [DONE]
Allow ".ddro" file extensions in addition to ".diderot"
Be able to output values of type tensor[2,2] and tensor[3,3];
(currently only scalars & vectors). Want to add some regression tests
based on this and currently can't
[GLK:1] Add a clamp function, which takes three arguments; either
three scalars:
clamp(lo, hi, x) = max(lo, min(hi, x))
or three vectors of the same size:
clamp(lo, hi, [x,y]) = [max(lo[0], min(hi[0], x)),
max(lo[1], min(hi[1], y))]
This would be useful in many current Diderot programs.
One question: clamp(x, lo, hi) is the argument order used in OpenCL
and other places, but clamp(lo, hi, x) is much more consistent with
lerp(lo, hi, x), hence GLK's preference
[DONE]
[GLK:2] Proper handling of stabilize method
allow "*" to represent "modulate": per-component multiplication of
vectors, and vectors only (not tensors of order 2 or higher). Once
sequences are implemented this should be removed: the operation is not
invariant WRT basis so it is not a legit vector computation.
implicit type promotion of integers to reals where reals are
required (e.g. not exponentiation "^")
[GLK:5] Save Diderot output to nrrd, instead of "mip.txt"
For grid of strands, save to similarly-shaped array
For list of strands, save to long 1-D (or 2-D for non-scalar output) list
For ragged things (like tractography output), will need to save both
complete list of values, as well as list of start indices and lengths
to index into complete list
[GLK:6] Use of Teem's "hest" command-line parser for getting
any "input" variables that are not defined in the source file.
[GLK:7] ability to declare a field so that probe positions are
*always* "inside"; with various ways of mapping the known image values
to non-existant index locations. One possible syntax emphasizes that
there is a index mapping function that logically precedes convolution:
F = bspln3 ⊛ (img ◦ clamp)
F = bspln3 ⊛ (img ◦ repeat)
F = bspln3 ⊛ (img ◦ mirror)
where "◦" or "∘" is used to indicate function composition
Level of differentiability in field type should be statement about how
much differentiation the program *needs*, rather than what the kernel
*provides*. The needed differentiability can be less than or equal to
the provided differentiability.
Use ∇⊗ etc. syntax
syntax [DONE]
typechecking
IL and codegen
Add type aliases for color types
rgb = real{3}
rgba = real{4}
==============================
MEDIUM TERM ================== (*needed* for particles)
==============================
run-time birth of strands
"initially" supports lists
"initially" supports lists of positions output from different
initalization Diderot program (or output from the same program;
e.g. using output of iso2d.diderot for one isovalue to seed the input
to another invocation of the same program)
Communication between strands: they have to be able to learn each
other's state (at the previous iteration). Early version of this can
have the network of neighbors be completely static (for running one
strand/pixel image computations). Later version with strands moving
through the domain will require some spatial data structure to
optimize discovery of neighbors.
============================
MEDIUM-ISH TERM ============ (to make Diderot more useful/effective)
============================
Python/ctypes interface to run-time
support for Python interop and GUI
Allow integer exponentiation ("^2") to apply to square matrices,
to represent repeated matrix multiplication
Alow X *= Y, X /= Y, X += Y, X -= Y to mean what they do in C,
provided that X*Y, X/Y, X+Y, X-Y are already supported.
Nearly every Diderot program would be simplified by this.
[DONE]
Put small 1-D and 2-D fields, when reconstructed specifically by tent
and when differentiation is not needed, into faster texture buffers.
test/illust-vr.diderot is good example of program that uses multiple
such 1-D fields basically as lookup-table-based function evaluation
expand trace in mid to low translation
extend norm (|exp|) to all tensor types [DONE for vectors and matrices]
determinant ("det") for tensor[3,3]
add ":" for tensor dot product (contracts out two indices
instead of one like •), valid for all pairs of tensors with
at least two indices
test/uninit.diderot:
documents need for better compiler error messages when output variables
are not initialized; the current messages are very cryptic
want: warnings when "D" (reserved for differentiation) is declared as
a variable name (get confusing error messages now)
==============================
LONG TERM ==================== (make Diderot more interesting/attractive from
============================== a research standpoint)
IL support for higher-order tensor values (matrices, etc).
tensor construction [DONE]
tensor indexing [DONE]
tensor slicing
verify that hessians work correctly [DONE]
Better handling of variables that determines the scope of a variable
based on its actual use, instead of where the user defined it. So,
for example, we should lift strand-invariant variables to global
scope. Also prune out useless variables, which should include field
variables after the translation to mid-il.
test/vr-kcomp2.diderot: Add support for code like
(F1 if x else F2)@pos
This will require duplication of the continuation of the conditional
(but we should only duplicate over the live-range of the result of the
conditional.
[GLK:8] Want: non-trivial field expressions & functions.
scalar fields from scalar fields F and G:
field#0(2)[] X = (sin(F) + 1.0)/2;
field#0(2)[] X = F*G;
scalar field of vector field magnitude:
image(2)[2] Vimg = load(...);
field#0(2)[] Vlen = |Vimg ⊛ bspln3|;
field of normalized vectors (for LIC and vector field feature extraction)
field#2(2)[2] F = ...
field#0(2)[2] V = normalize(F);
scalar field of gradient magnitude (for edge detection))
field#2(2)[] F = Fimg ⊛ bspln3;
field#0(2)[] Gmag = |∇F|;
scalar field of squared gradient magnitude (simpler to differentiate):
field#2(2)[] F = Fimg ⊛ bspln3;
field#0(2)[] Gmsq = ∇F•∇F;
There is value in having these, even if the differentiation of them is
not supported (hence the indication of "field#0" for these above)
Introduce region types (syntax region(d), where d is the dimension of the
region. One useful operator would be
dom : field#k(d)[s] -> region(d)
Then the inside test could be written as
pos ∈ dom(F)
We could further extend this approach to allow geometric definitions of
regions. It might also be useful to do inside tests in world space,
instead of image space.
co- vs contra- index distinction
Permit field composition:
field#2(3)[3] warp = bspln3 ⊛ warpData;
field#2(3)[] F = bspln3 ⊛ img;
field#2(3)[] Fwarp = F ◦ warp;
So Fwarp(x) = F(warp(X)). Chain rule can be used for differentation.
This will be instrumental for expressing non-rigid registration
methods (but those will require co-vs-contra index distinction)
Allow the convolution to be specified either as a single 1D kernel
(as we have it now):
field#2(3)[] F = bspln3 ⊛ img;
or, as a tensor product of kernels, one for each axis, e.g.
field#0(3)[] F = (bspln3 ⊗ bspln3 ⊗ tent) ⊛ img;
This is especially important for things like time-varying fields
and the use of scale-space in field visualization: one axis of the
must be convolved with a different kernel during probing.
What is very unclear is how, in such cases, we should notate the
gradient, when we only want to differentiate with respect to some
subset of the axes. One ambitious idea would be:
field#0(3)[] Ft = (bspln3 ⊗ bspln3 ⊗ tent) ⊛ img; // 2D time-varying field
field#0(2)[] F = lambda([x,y], Ft([x,y,42.0])) // restriction to time=42.0
vec2 grad = ∇F([x,y]); // 2D gradient
Tensors of order 3 (e.g. gradients of diffusion tensor fields, or
hessians of vector fields) and order 4 (e.g. Hessians of diffusion
tensor fields).
representation of tensor symmetry
(have to identify the group of index permutations that are symmetries)
dot works on all tensors
outer works on all tensors
Help for debugging Diderot programs: need to be able to uniquely
identify strands, and for particular strands that are known to behave
badly, do something like printf or other logging of their computations
and updates.
Permit writing dimensionally general code: Have some statement of the
dimension of the world "W" (or have it be learned from one particular
field of interest), and then able to write "vec" instead of
"vec2/vec3", and perhaps "tensor[W,W]" instead of
"tensor[2,2]/tensor[3,3]"
Traits: all things things that have boilerplate code (especially
volume rendering) should be expressed in terms of the unique
computational core. Different kinds of streamline/tractography
computation will be another example, as well as particle systems.
Einstein summation notation
"tensor comprehension" (like list comprehension)
Fields coming from different sources of data:
* triangular or tetrahedral meshes over 2D or 3D domains (of the
source produced by finite-element codes; these will come with their
own specialized kinds of reconstruction kernels, called "basis
functions" in this context)
* Large point clouds, with some radial basis function around each point,
which will be tuned by parameters of the point (at least one parameter
giving some notion of radius)
======================
BUGS =================
======================
test/zslice2.diderot:
// HEY (bug) bspln5 leads to problems ...
// uncaught exception Size [size]
// raised at c-target/c-target.sml:47.15-47.19
//field#4(3)[] F = img ⊛ bspln5;