Home My Page Projects Code Snippets Project Openings diderot
Summary Activity Tracker Tasks SCM

SCM Repository

[diderot] Diff of /trunk/TODO
ViewVC logotype

Diff of /trunk/TODO

Parent Directory Parent Directory | Revision Log Revision Log | View Patch Patch

revision 1155, Sun May 8 14:43:30 2011 UTC revision 1295, Thu Jun 9 06:42:54 2011 UTC
# Line 1  Line 1 
1  NOTE: GLK's approximate ranking of 5 most important tagged with  NOTE: GLK's approximate ranking of 8 most important tagged with
2  [GLK:1], [GLK:2], ...  [GLK:1], [GLK:2], ...
3    
4  ==============================  ========================
5  other SHORT TERM =============  (including needed for LIC)  SHORT TERM ============= (*needed* for streamlines & tractography)
6  ==============================  ========================
   
 Add a clamp function, which takes three arguments; either three scalars:  
   clamp(x, minval, maxval)  = max(minval, min(maxval, x))  
 or three vectors of the same size:  
   clamp([x,y], minvec, maxvec)  = [max(minvec[0], min(maxvec[0], x)),  
                                    max(minvec[1], min(maxvec[1], y))]  
 This would be useful in many current Diderot programs.  
 One question: clamp(x, minval, maxval) is the argument order  
 used in OpenCL and other places, but clamp(minval, maxval, x)  
 would be more consistent with lerp(minout, maxout, x).  
7    
8  Level of differentiability in field type should be statement about how  Remove CL from compiler
 much differentiation the program *needs*, rather than what the kernel  
 *provides*.  The needed differentiability can be less than or equal to  
 the provided differentiability.  
9    
10  [GLK:1] Add sequence types (needed for evals & evecs)  [GLK:3] Add sequence types (needed for evals & evecs)
11      syntax      syntax
12          types: ty '{' INT '}'          types: ty '{' INT '}'
13          value construction: '{' e1 ',' … ',' en '}'          value construction: '{' e1 ',' … ',' en '}'
14          indexing: e '{' e '}'          indexing: e '{' e '}'
15    
16  IL support for higher-order tensor values (matrices, etc).  [GLK:4] evals & evecs for symmetric tensor[2,2] and
17      tensor construction [DONE]  tensor[3,3] (requires sequences)
     tensor indexing [DONE]  
     tensor slicing  
     verify that hessians work correctly [DONE]  
18    
19  Use ∇⊗ etc. syntax  ability to emit/track/record variables into dynamically re-sized
20      syntax [DONE]  runtime buffer
     typechecking  
     IL and codegen  
21    
22  test/uninit.diderot:  tensor fields: convolution on general tensor images
 documents need for better compiler error messages when output variables  
 are not initialized; the current messages are very cryptic  
23    
24  determinant ("det") for tensor[3,3]  ========================
25    SHORT-ISH TERM ========= (to make using Diderot less annoying to
26    ========================  program in, and slow to execute)
27    
28  expand trace in mid to low translation  value-numbering optimization [DONE]
29    
30  value-numbering optimization  Allow ".ddro" file extensions in addition to ".diderot"
31    
32  Add type aliases for color types  Be able to output values of type tensor[2,2] and tensor[3,3];
33      rgb = real{3}  (currently only scalars & vectors).  Want to add some regression tests
34      rgba = real{4}  based on this and currently can't
35    
36  ==============================  [GLK:1] Add a clamp function, which takes three arguments; either
37  MEDIUM TERM ================== (including needed for streamlines & tractography)  three scalars:
38  ==============================    clamp(lo, hi, x)  = max(lo, min(hi, x))
39    or three vectors of the same size:
40      clamp(lo, hi, [x,y])  = [max(lo[0], min(hi[0], x)),
41                               max(lo[1], min(hi[1], y))]
42    This would be useful in many current Diderot programs.
43    One question: clamp(x, lo, hi) is the argument order used in OpenCL
44    and other places, but clamp(lo, hi, x) is much more consistent with
45    lerp(lo, hi, x), hence GLK's preference
46    [DONE]
47    
48    [GLK:2] Proper handling of stabilize method
49    
50    allow "*" to represent "modulate": per-component multiplication of
51    vectors, and vectors only (not tensors of order 2 or higher).  Once
52    sequences are implemented this should be removed: the operation is not
53    invariant WRT basis so it is not a legit vector computation.
54    
55  [GLK:1] evals & evecs for symmetric tensor[3,3] (requires sequences)  implicit type promotion of integers to reals where reals are
56    required (e.g. not exponentiation "^")
57    
58  [GLK:2] Save Diderot output to nrrd, instead of "mip.txt"  [GLK:5] Save Diderot output to nrrd, instead of "mip.txt"
59    For grid of strands, save to similarly-shaped array    For grid of strands, save to similarly-shaped array
60    For list of strands, save to long 1-D (or 2-D for non-scalar output) list    For list of strands, save to long 1-D (or 2-D for non-scalar output) list
61    For ragged things (like tractography output), will need to save both    For ragged things (like tractography output), will need to save both
62      complete list of values, as well as list of start indices and lengths      complete list of values, as well as list of start indices and lengths
63      to index into complete list      to index into complete list
64    
65  [GLK:3] Use of Teem's "hest" command-line parser for getting  [GLK:6] Use of Teem's "hest" command-line parser for getting
66  any input variables that are not defined in the source file  any "input" variables that are not defined in the source file.
67    
68  [GLK:4] ability to declare a field so that probe positions are  [GLK:7] ability to declare a field so that probe positions are
69  *always* "inside"; with various ways of mapping the known image values  *always* "inside"; with various ways of mapping the known image values
70  to non-existant index locations.  One possible syntax emphasizes that  to non-existant index locations.  One possible syntax emphasizes that
71  there is a index mapping function that logically precedes convolution:  there is a index mapping function that logically precedes convolution:
# Line 76  Line 74 
74    F = bspln3 ⊛ (img ◦ mirror)    F = bspln3 ⊛ (img ◦ mirror)
75  where "◦" or "∘" is used to indicate function composition  where "◦" or "∘" is used to indicate function composition
76    
77  extend norm (|exp|) to all tensor types [DONE for vectors and matrices]  Level of differentiability in field type should be statement about how
78    much differentiation the program *needs*, rather than what the kernel
79  ability to emit/track/record variables into dynamically re-sized  *provides*.  The needed differentiability can be less than or equal to
80  runtime buffer  the provided differentiability.
   
 Want: allow 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.  
81    
82  [GLK:5] Want: non-trivial field expressions & functions:  Use ∇⊗ etc. syntax
83    image(2)[2] Vimg = load(...);      syntax [DONE]
84    field#0(2)[] Vlen = |Vimg ⊛ bspln3|;      typechecking
85  to get a scalar field of vector length, or      IL and codegen
   field#2(2)[] F = Fimg ⊛ bspln3;  
   field#0(2)[] Gmag = |∇F|;  
 to get a scalar field of gradient magnitude, or  
   field#2(2)[] F = Fimg ⊛ bspln3;  
   field#0(2)[] Gmsq = ∇F•∇F;  
 to get a scalar field of squared gradient magnitude, which is simpler  
 to differentiate.  However, there is value in having these, even if  
 the differentiation of them is not supported (hence the indication  
 of "field#0" for these above)  
   
 Want: ability to apply "normalize" to a field itself, e.g.  
   field#0(2)[2] V = normalize(Vimg ⊛ ctmr);  
 so that V(x) = normalize((Vimg ⊛ ctmr)(x)).  
 Having this would simplify expression of standard LIC method, and  
 would also help express other vector field expressions that arise  
 in vector field feature exraction.  
86    
87  tensor fields: convolution on general tensor images  Add type aliases for color types
88        rgb = real{3}
89        rgba = real{4}
90    
91  ==============================  ==============================
92  other MEDIUM TERM ============ (needed for particles)  MEDIUM TERM ================== (*needed* for particles)
93  ==============================  ==============================
94    
 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  
   
95  run-time birth of strands  run-time birth of strands
96    
97  "initially" supports lists  "initially" supports lists
98    
99  "initially" supports lists of positions output from  "initially" supports lists of positions output from different
100  different initalization Diderot program  initalization Diderot program (or output from the same program;
101    e.g. using output of iso2d.diderot for one isovalue to seed the input
102    to another invocation of the same program)
103    
104    Communication between strands: they have to be able to learn each
105    other's state (at the previous iteration).  Early version of this can
106    have the network of neighbors be completely static (for running one
107    strand/pixel image computations).  Later version with strands moving
108    through the domain will require some spatial data structure to
109    optimize discovery of neighbors.
110    
111    ============================
112    MEDIUM-ISH TERM ============ (to make Diderot more useful/effective)
113    ============================
114    
115    Python/ctypes interface to run-time
116    
117  spatial data structure that permits strands' queries of neighbors  support for Python interop and GUI
118    
119  proper handling of stabilize method  Allow integer exponentiation ("^2") to apply to square matrices,
120    to represent repeated matrix multiplication
121    
122  test/vr-kcomp2.diderot: Add support for code like  Alow X *= Y, X /= Y, X += Y, X -= Y to mean what they do in C,
123    provided that X*Y, X/Y, X+Y, X-Y are already supported.
124    Nearly every Diderot program would be simplified by this.
125    
126          (F1 if x else F2)@pos  Put small 1-D and 2-D fields, when reconstructed specifically by tent
127    and when differentiation is not needed, into faster texture buffers.
128    test/illust-vr.diderot is good example of program that uses multiple
129    such 1-D fields basically as lookup-table-based function evaluation
130    
131  This will require duplication of the continuation of the conditional  expand trace in mid to low translation
132  (but we should only duplicate over the live-range of the result of the  
133  conditional.  extend norm (|exp|) to all tensor types [DONE for vectors and matrices]
134    
135    determinant ("det") for tensor[3,3]
136    
137  add ":" for tensor dot product (contracts out two indices  add ":" for tensor dot product (contracts out two indices
138  instead of one like •), valid for all pairs of tensors with  instead of one like •), valid for all pairs of tensors with
139  at least two indices  at least two indices
140    
141  ==============================  test/uninit.diderot:
142  other MEDIUM TERM ============  documents need for better compiler error messages when output variables
143  ==============================  are not initialized; the current messages are very cryptic
144    
145  want: warnings when "D" (reserved for differentiation) is declared as  want: warnings when "D" (reserved for differentiation) is declared as
146  a variable name (get confusing error messages now)  a variable name (get confusing error messages now)
147    
 support for Python interop and GUI  
   
 Python/ctypes interface to run-time  
   
 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 data, or  
 other multi-dimensional fields where one axis of the domain is very  
 different from the rest.  What is very unclear is how, in such cases,  
 we should notate the gradient, when we only want to differentiate with  
 respect to some of the axes.  
   
 ==============================  
 LONG TERM ====================  
148  ==============================  ==============================
149    LONG TERM ==================== (make Diderot more interesting/attractive from
150    ==============================  a research standpoint)
151    
152    IL support for higher-order tensor values (matrices, etc).
153        tensor construction [DONE]
154        tensor indexing [DONE]
155        tensor slicing
156        verify that hessians work correctly [DONE]
157    
158  Better handling of variables that determines the scope of a variable  Better handling of variables that determines the scope of a variable
159  based on its actual use, instead of where the user defined it.  So,  based on its actual use, instead of where the user defined it.  So,
# Line 172  Line 161 
161  scope.  Also prune out useless variables, which should include field  scope.  Also prune out useless variables, which should include field
162  variables after the translation to mid-il.  variables after the translation to mid-il.
163    
164    test/vr-kcomp2.diderot: Add support for code like
165            (F1 if x else F2)@pos
166    This will require duplication of the continuation of the conditional
167    (but we should only duplicate over the live-range of the result of the
168    conditional.
169    
170    [GLK:8] Want: non-trivial field expressions & functions.
171    scalar fields from scalar fields F and G:
172      field#0(2)[] X = (sin(F) + 1.0)/2;
173      field#0(2)[] X = F*G;
174    scalar field of vector field magnitude:
175      image(2)[2] Vimg = load(...);
176      field#0(2)[] Vlen = |Vimg ⊛ bspln3|;
177    field of normalized vectors (for LIC and vector field feature extraction)
178      field#2(2)[2] F = ...
179      field#0(2)[2] V = normalize(F);
180    scalar field of gradient magnitude (for edge detection))
181      field#2(2)[] F = Fimg ⊛ bspln3;
182      field#0(2)[] Gmag = |∇F|;
183    scalar field of squared gradient magnitude (simpler to differentiate):
184      field#2(2)[] F = Fimg ⊛ bspln3;
185      field#0(2)[] Gmsq = ∇F•∇F;
186    There is value in having these, even if the differentiation of them is
187    not supported (hence the indication of "field#0" for these above)
188    
189    Introduce region types (syntax region(d), where d is the dimension of the
190    region.  One useful operator would be
191            dom : field#k(d)[s] -> region(d)
192    Then the inside test could be written as
193            pos ∈ dom(F)
194    We could further extend this approach to allow geometric definitions of
195    regions.  It might also be useful to do inside tests in world space,
196    instead of image space.
197    
198  co- vs contra- index distinction  co- vs contra- index distinction
199    
200  some indication of tensor symmetry  Permit field composition:
201      field#2(3)[3] warp = bspln3 ⊛ warpData;
202      field#2(3)[] F = bspln3 ⊛ img;
203      field#2(3)[] Fwarp = F ◦ warp;
204    So Fwarp(x) = F(warp(X)).  Chain rule can be used for differentation.
205    This will be instrumental for expressing non-rigid registration
206    methods (but those will require co-vs-contra index distinction)
207    
208    Allow the convolution to be specified either as a single 1D kernel
209    (as we have it now):
210      field#2(3)[] F = bspln3 ⊛ img;
211    or, as a tensor product of kernels, one for each axis, e.g.
212      field#0(3)[] F = (bspln3 ⊗ bspln3 ⊗ tent) ⊛ img;
213    This is especially important for things like time-varying fields
214    and the use of scale-space in field visualization: one axis of the
215    must be convolved with a different kernel during probing.
216    What is very unclear is how, in such cases, we should notate the
217    gradient, when we only want to differentiate with respect to some
218    subset of the axes.  One ambitious idea would be:
219      field#0(3)[] Ft = (bspln3 ⊗ bspln3 ⊗ tent) ⊛ img; // 2D time-varying field
220      field#0(2)[] F = lambda([x,y], Ft([x,y,42.0]))     // restriction to time=42.0
221      vec2 grad = ∇F([x,y]);                             // 2D gradient
222    
223    Tensors of order 3 (e.g. gradients of diffusion tensor fields, or
224    hessians of vector fields) and order 4 (e.g. Hessians of diffusion
225    tensor fields).
226    
227    representation of tensor symmetry
228  (have to identify the group of index permutations that are symmetries)  (have to identify the group of index permutations that are symmetries)
229    
230  dot works on all tensors  dot works on all tensors
231    
232  outer works on all tensors  outer works on all tensors
233    
234    Help for debugging Diderot programs: need to be able to uniquely
235    identify strands, and for particular strands that are known to behave
236    badly, do something like printf or other logging of their computations
237    and updates.
238    
239    Permit writing dimensionally general code: Have some statement of the
240    dimension of the world "W" (or have it be learned from one particular
241    field of interest), and then able to write "vec" instead of
242    "vec2/vec3", and perhaps "tensor[W,W]" instead of
243    "tensor[2,2]/tensor[3,3]"
244    
245    Traits: all things things that have boilerplate code (especially
246    volume rendering) should be expressed in terms of the unique
247    computational core.  Different kinds of streamline/tractography
248    computation will be another example, as well as particle systems.
249    
250  Einstein summation notation  Einstein summation notation
251    
252  "tensor comprehension" (like list comprehension)  "tensor comprehension" (like list comprehension)
253    
254    Fields coming from different sources of data:
255    * triangular or tetrahedral meshes over 2D or 3D domains (of the
256      source produced by finite-element codes; these will come with their
257      own specialized kinds of reconstruction kernels, called "basis
258      functions" in this context)
259    * Large point clouds, with some radial basis function around each point,
260      which will be tuned by parameters of the point (at least one parameter
261      giving some notion of radius)
262    
263  ======================  ======================
264  BUGS =================  BUGS =================
265  ======================  ======================

Legend:
Removed from v.1155  
changed lines
  Added in v.1295

root@smlnj-gforge.cs.uchicago.edu
ViewVC Help
Powered by ViewVC 1.0.0