8a78217 | Barak A. Pearlmutter | 22 February 2018, 11:14:43 UTC | terminal newline | 22 February 2018, 11:14:43 UTC |
7df5281 | Sophia Gold | 22 February 2018, 09:11:18 UTC | .sc = Scheme | 22 February 2018, 09:11:18 UTC |
4032d37 | Barak A. Pearlmutter | 30 September 2016, 10:10:18 UTC | update README.md | 30 September 2016, 10:10:18 UTC |
2092ecc | Barak A. Pearlmutter | 20 November 2014, 09:59:09 UTC | add tools/README, direct unbuff users to /usr/bin/stdbuff | 29 September 2016, 16:18:52 UTC |
17b543a | Barak A. Pearlmutter | 06 June 2014, 11:36:07 UTC | autogen.sh | 29 September 2016, 16:18:52 UTC |
d6abcbc | Barak A. Pearlmutter | 06 June 2014, 11:19:17 UTC | update git ignorance | 29 September 2016, 16:18:52 UTC |
0a09aa6 | Barak A. Pearlmutter | 06 June 2014, 11:12:10 UTC | remove qobischeme from autotools goo - in its own repo now | 29 September 2016, 16:18:52 UTC |
050b930 | Barak A. Pearlmutter | 21 November 2012, 22:07:43 UTC | compile executable tools/sum | 29 September 2016, 16:18:52 UTC |
e9188e6 | Barak A. Pearlmutter | 21 November 2012, 19:16:22 UTC | reduce warnings in mplayerwin.c | 29 September 2016, 16:18:52 UTC |
69e0b5f | Barak A. Pearlmutter | 21 November 2012, 19:16:06 UTC | git ignore tools/ executables | 29 September 2016, 16:18:52 UTC |
d2106fb | Barak A. Pearlmutter | 21 November 2012, 19:15:11 UTC | make mplayerwin and unbuff | 29 September 2016, 16:18:52 UTC |
7559fa1 | Barak A. Pearlmutter | 21 November 2012, 18:58:16 UTC | mplayerwin.c from jeff via email | 29 September 2016, 16:18:52 UTC |
97ab4c9 | Barak A. Pearlmutter | 01 March 2012, 19:24:47 UTC | autoupdate | 29 September 2016, 16:18:52 UTC |
ad5d137 | Alexey Radul | 08 December 2011, 13:29:46 UTC | Proper tangent solves the curried plus problem. | 29 September 2016, 16:18:52 UTC |
7b716f4 | Alexey Radul | 07 December 2011, 12:20:53 UTC | I should put my terrible hacks in properly. | 29 September 2016, 16:18:52 UTC |
004ac3a | Alexey Radul | 07 December 2011, 12:19:57 UTC | Standard options suggested by Jeff. | 29 September 2016, 16:18:52 UTC |
58d9d04 | Alexey Radul | 05 December 2011, 21:48:04 UTC | Another fun example to think about. | 29 September 2016, 16:18:52 UTC |
3b56dc1 | Alexey Radul | 02 December 2011, 23:17:05 UTC | A little abstraction to make running the benchmarks against DVL a little easier. | 29 September 2016, 16:18:52 UTC |
b946675 | Alexey Radul | 29 November 2011, 03:57:43 UTC | Raising to a given bundle level fixes DVL analysis errors on saddle.vlad. I am somewhat annoyed that it appeared to be necessary to do this in so many places; on the other hand, I am quite pleased that the DVL debugging facilities (such as they are) allowed me to identify the appropriate spots with relative ease. | 29 September 2016, 16:18:52 UTC |
226feb4 | Alexey Radul | 29 November 2011, 01:57:07 UTC | Now that I named this abstraction, I would use it. | 29 September 2016, 16:18:52 UTC |
a9ae899 | Alexey Radul | 29 November 2011, 01:48:47 UTC | Work around a limitation of DVL. This should not to affect Stalingrad's understanding of this program in the least. Yes, we need to come up with a real solution, but while the benchmark deadline looms I want to push quick hacks as far as they will go. | 29 September 2016, 16:18:52 UTC |
c9f5221 | Alexey Radul | 16 November 2011, 17:46:40 UTC | Oops, didn't mean to comment out all the tests. | 29 September 2016, 16:18:52 UTC |
b637121 | Alexey Radul | 15 November 2011, 18:09:36 UTC | Work around a known bug in DVL. Namely, that DVL doesn't like quoted lists with floating point numbers. | 29 September 2016, 16:18:52 UTC |
b206013 | Oleksandr Manzyuk | 27 October 2011, 15:30:49 UTC | Specialize these polymorphic functions. This gives another 2.5x speedup on both examples (0.126 vs. 0.296 for the particle example and 0.08 vs. 0.216 for the saddle example). The total speedup compared with the original unoptimized version compiled without -O flag is 20x. | 29 September 2016, 16:18:52 UTC |
a1b5f90 | Oleksandr Manzyuk | 27 October 2011, 15:27:16 UTC | Add top-level type signatures. I hate when they are omitted. | 29 September 2016, 16:18:52 UTC |
da44b5e | Oleksandr Manzyuk | 27 October 2011, 15:25:51 UTC | Replace 'error "unimplemented"' with actual implementations. | 29 September 2016, 16:18:52 UTC |
202ed37 | Oleksandr Manzyuk | 27 October 2011, 15:23:56 UTC | Don't need this extension anymore. Surprisingly, making arithmetic strict doesn't improve the performance (and sometimes make it worse). | 29 September 2016, 16:18:52 UTC |
bb11889 | Oleksandr Manzyuk | 27 October 2011, 15:22:29 UTC | Don't need this bang pattern. | 29 September 2016, 16:18:52 UTC |
a18ab33 | Oleksandr Manzyuk | 27 October 2011, 14:06:20 UTC | Replace explicit lambda with function composition. | 29 September 2016, 16:18:52 UTC |
a12bbe8 | Oleksandr Manzyuk | 27 October 2011, 14:04:45 UTC | Oh, man. Replace 'foldl (+) 0' with 'sum'. 'foldl (+) 0' is leaky, the proper version is 'foldl' (+) 0' and it is called 'sum'. Speeds up the particle example by a factor of 2. | 29 September 2016, 16:18:52 UTC |
a3d72f9 | Oleksandr Manzyuk | 27 October 2011, 13:27:32 UTC | Make 'lift' strict and inlinable. This seems to marginally speedup both programs. | 29 September 2016, 16:18:52 UTC |
62b9fc7 | Oleksandr Manzyuk | 27 October 2011, 13:15:38 UTC | Make Bundle strict. This speeds up the particle example by a factor of 4. The saddle example becomes 20% faster. (For the saddle example, the difference between compiling the program with and without -O flag is more impressive: 0.277 vs. 1.649; for the particle example, this ratio is less than 2). | 29 September 2016, 16:18:52 UTC |
58dcffd | Oleksandr Manzyuk | 27 October 2011, 11:37:48 UTC | Get rid of that n+1 pattern. | 29 September 2016, 16:18:52 UTC |
bdcfb0c | Alexey Radul | 09 September 2011, 18:31:32 UTC | The motivation for these stream experiments. | 29 September 2016, 16:18:52 UTC |
e4ca33a | Alexey Radul | 20 June 2011, 12:32:37 UTC | More thoughts about loop fusion. | 29 September 2016, 16:18:52 UTC |
25d9e6f | Alexey Radul | 20 June 2011, 11:30:17 UTC | Some commentary on variations on the theme. | 29 September 2016, 16:18:52 UTC |
642d38a | Alexey Radul | 20 June 2011, 10:56:50 UTC | Loop fusion. It seems to work. | 29 September 2016, 16:18:52 UTC |
b9bfc31 | Alexey Radul | 15 June 2011, 12:07:38 UTC | A little loop fusion experiment, just to check. | 29 September 2016, 16:18:52 UTC |
8ddf698 | Alexey Radul | 27 May 2011, 18:16:27 UTC | A file of random stuff where I was experimenting with Stalingrad's behavior. I don't actually know what to make of it. The latter examples observe that Stalingrad does not zero out constants in procedure bodies. | 29 September 2016, 16:14:38 UTC |
c22aad8 | Alexey Radul | 21 May 2011, 20:23:29 UTC | Updating the pointer to the BCL-AD implementations. | 29 September 2016, 16:14:38 UTC |
a8ef60c | Alexey Radul | 19 May 2011, 16:59:20 UTC | Ha! I discovered why this example fails in VL and DVL. | 29 September 2016, 16:14:38 UTC |
8c91603 | Alexey Radul | 31 March 2011, 16:30:07 UTC | Specifying order of evaluation. | 29 September 2016, 16:14:38 UTC |
d881478 | Alexey Radul | 31 March 2011, 16:26:47 UTC | Inserting an algebraic noop for the sake of the DVL type inference (which fails to unify a real with a bundle). | 29 September 2016, 16:14:38 UTC |
7e5c3f2 | Alexey Radul | 11 March 2011, 20:41:27 UTC | Tabs-B-Gone. | 29 September 2016, 16:14:38 UTC |
738df5d | Alexey Radul | 11 March 2011, 20:20:05 UTC | Sasha found a bug and convinced me that it was a real bug. Tests changed. | 29 September 2016, 16:14:38 UTC |
6067a14 | Alexey Radul | 11 March 2011, 14:07:45 UTC | That really needs to be sure and be a tab character. | 29 September 2016, 16:14:38 UTC |
84f4929 | Alexey Radul | 03 March 2011, 17:44:04 UTC | Trying to test confluence. Unfortunately, Stalingrad refuses (I think it doesn't like the constants in closure bodies) and SLAD fails. | 29 September 2016, 16:14:38 UTC |
a944691 | Alexey Radul | 03 March 2011, 17:30:32 UTC | Another VLAD implementation is now approximately online. | 29 September 2016, 16:14:38 UTC |
77d0fc8 | Alexey Radul | 03 March 2011, 17:28:24 UTC | Abstracting away the differences between SLAD and VL as VLAD implementations (my, how minor they are!) | 29 September 2016, 16:14:38 UTC |
ef02f95 | Alexey Radul | 03 March 2011, 17:15:48 UTC | Easier to comment these files out now. | 29 September 2016, 16:14:38 UTC |
9a49254 | Alexey Radul | 28 February 2011, 16:22:14 UTC | Poking at Stalingrad's ability (or lack thereof) to deal with derivatives of functions that emit structures, whose primal answers are known at analysis time. | 29 September 2016, 16:14:38 UTC |
99669f0 | Alexey Radul | 28 February 2011, 16:21:48 UTC | Another nice example of perturbations that should not be confused. | 29 September 2016, 16:14:38 UTC |
a4109ed | Alexey Radul | 26 February 2011, 22:09:11 UTC | Hm. I ran Stalingrad, and now it gets a different answer than it did before. What's going on? Did I mess something up by abstracting the gradient computations? Do different machines produce different answers? What's going on here? | 29 September 2016, 16:14:38 UTC |
885d6ea | Alexey Radul | 26 February 2011, 15:00:18 UTC | Unknown number of derivative levels. The compiler takes longer than 10 minutes on this on khazad-dum. Is it in an infinite loop? | 29 September 2016, 16:14:38 UTC |
fa89182 | Alexey Radul | 23 February 2011, 01:16:38 UTC | Nothing is gained and many CPU cycles are lost by running all the long-running tests to 50 minutes. I propose cutting them off at 10. | 29 September 2016, 16:14:38 UTC |
baeb6d5 | Alexey Radul | 23 February 2011, 00:11:27 UTC | If I want to actually test the compiler properly (as opposed to, say, the quality of the compiler's error messages), I should probably try and defeat the abstract analysis in the places where I expect it to be defeated. | 29 September 2016, 16:14:38 UTC |
eec95be | Alexey Radul | 23 February 2011, 00:01:44 UTC | Ran particle in SLAD again. I am very surprised -- am I confusing the perturbations in a different way now? | 29 September 2016, 16:14:38 UTC |
6e34cec | Alexey Radul | 22 February 2011, 23:47:19 UTC | I find myself wanting to do the test-preparation actions often enough that there should probably be a task for them. | 29 September 2016, 16:14:38 UTC |
fee36cb | Alexey Radul | 22 February 2011, 23:40:32 UTC | Simplify, relying on the automatic addition of structure-extraction code. | 29 September 2016, 16:14:38 UTC |
280e7d9 | Alexey Radul | 22 February 2011, 23:30:24 UTC | Fold the if into the cond. | 29 September 2016, 16:14:38 UTC |
d8200c1 | Alexey Radul | 22 February 2011, 23:26:29 UTC | Moving around and explaining the testing code that now does a better job with compiling tests that produce structured outputs. | 29 September 2016, 16:14:38 UTC |
ca299b6 | Alexey Radul | 22 February 2011, 22:10:09 UTC | I want those bogons to really stand out. | 29 September 2016, 16:14:38 UTC |
4ee8152 | Alexey Radul | 22 February 2011, 22:00:52 UTC | I need union-free *and* to actually force the program to compute the answer I'm looking for, instead of having the analysis assume it knows the answer. | 29 September 2016, 16:14:38 UTC |
a19d9c7 | Alexey Radul | 22 February 2011, 21:51:37 UTC | Trying to support booleans in a union-free manner. | 29 September 2016, 16:14:38 UTC |
e2300e1 | Alexey Radul | 22 February 2011, 21:44:30 UTC | Supporting booleans in the careful writing hack. | 29 September 2016, 16:14:38 UTC |
c993669 | Alexey Radul | 22 February 2011, 21:32:38 UTC | Right. I should get right the case where the original answer was multiform. | 29 September 2016, 16:14:38 UTC |
23709e5 | Alexey Radul | 22 February 2011, 21:24:28 UTC | Do not produce error messages when trying to inspect parsed but not specialized expectations. | 29 September 2016, 16:14:38 UTC |
4f53968 | Alexey Radul | 22 February 2011, 21:14:09 UTC | More precise reason for why a certain test is rejected. | 29 September 2016, 16:14:38 UTC |
4cfc2db | Alexey Radul | 22 February 2011, 21:12:07 UTC | Drafting a mechanism by which the compiler can be tested on examples that emit structured answers. The trick is that if I know the expected shape, I can write code into the testee program that will destructure exactly that shape and write the real numbers. | 29 September 2016, 16:14:38 UTC |
44ac913 | Alexey Radul | 22 February 2011, 18:47:33 UTC | Second-order derivative of function-returning-function. | 29 September 2016, 16:14:38 UTC |
63fda50 | Alexey Radul | 22 February 2011, 18:38:00 UTC | Another example with a data structure instead of a function. Right. I should remember that derivative-R expects the output to be the real numbers, so it doesn't apply to these new, funky examples. | 29 September 2016, 16:14:38 UTC |
07592c0 | Alexey Radul | 22 February 2011, 18:34:02 UTC | And its reverse-mode variant. | 29 September 2016, 16:14:38 UTC |
8492ddf | Alexey Radul | 22 February 2011, 18:32:16 UTC | Here's an example to test the universality of AD. | 29 September 2016, 16:14:38 UTC |
7eb4d2b | Alexey Radul | 18 February 2011, 16:43:27 UTC | Making sure that REAL? can deal with abstract reals. | 29 September 2016, 16:14:38 UTC |
5baac05 | Alexey Radul | 18 February 2011, 13:45:44 UTC | The executable is named "vl", not "run-vl". | 29 September 2016, 16:14:38 UTC |
6145b87 | Alexey Radul | 18 February 2011, 13:43:41 UTC | Now that SLAD and VL both implement (include "foo"), I don't need to detect "bad definitions" any more. | 29 September 2016, 16:14:38 UTC |
2df78cc | Alexey Radul | 18 February 2011, 13:38:27 UTC | If this is to remain a "fast" example, it should explicitly introduce imprecision (for example, for the benefit of implementations that do not support -imprecise-inexacts). | 29 September 2016, 16:14:37 UTC |
987cd86 | Alexey Radul | 18 February 2011, 13:03:00 UTC | Moving the type-testing examples out into their own file. The main purpose of this is to open out the multiform defining car and cdr without interfering with later definitions of car and cdr, so that I stop getting that annoying "bad definition" error. | 29 September 2016, 16:14:37 UTC |
5dd47aa | Alexey Radul | 18 February 2011, 12:51:29 UTC | Why am I even testing for quasiquote? | 29 September 2016, 16:14:37 UTC |
ab3eb87 | Alexey Radul | 17 February 2011, 17:10:10 UTC | Reverse-mode versions of all the basics. | 29 September 2016, 16:14:37 UTC |
5feadc1 | Alexey Radul | 17 February 2011, 17:05:11 UTC | Importing some basic examples from the SLAD sources. It seems that the last one tickles a form of perturbation confusion that does not occur in the other examples (at least not those that finish quickly). | 29 September 2016, 16:14:37 UTC |
cc409e1 | Alexey Radul | 17 February 2011, 13:55:48 UTC | The executable is now named slad, not run-slad. | 29 September 2016, 16:14:37 UTC |
59cd980 | Alexey Radul | 16 February 2011, 20:17:00 UTC | I think I want failure-summary to be the default output at the end of make rather than test-failures, to avoid cluttering the output with all the SLAD unbound variables. The real answer, though, is some sort of "expected failure" system, or to do a better job of pre-filtering the examples that SLAD will predictably fall down on. Is there a good way to do that in parallel instead of in the serial test suite preparation phase? | 29 September 2016, 16:14:37 UTC |
cd6009e | Alexey Radul | 16 February 2011, 20:16:24 UTC | tool/failure-summary is now grammatically correct when there are 0 unbound variable failures. | 29 September 2016, 16:14:37 UTC |
95cce02 | Alexey Radul | 16 February 2011, 15:43:55 UTC | Approximate equality of floating point results inside structured outputs too. | 29 September 2016, 16:14:37 UTC |
761e611 | Alexey Radul | 16 February 2011, 15:43:22 UTC | SLAD computed an answer for probabilistic prolog. | 29 September 2016, 16:14:37 UTC |
560c176 | Alexey Radul | 16 February 2011, 15:42:38 UTC | SLAD computed a different answer than the one I had there. I am concerned. | 29 September 2016, 16:14:37 UTC |
647bb94 | Alexey Radul | 15 February 2011, 16:47:10 UTC | SLAD computed another answer. | 29 September 2016, 16:14:37 UTC |
4fafb96 | Alexey Radul | 15 February 2011, 16:41:39 UTC | Hoisting derivative procedure constructions out of loops makes this code friendlier to expository interpreters. | 29 September 2016, 16:14:37 UTC |
281dc00 | Alexey Radul | 15 February 2011, 14:19:22 UTC | SLAD computed a candidate answer to probabilistic-lambda-calculus in forward mode (and, by extension, the same answer can be presumed for reverse mode). | 29 September 2016, 16:14:37 UTC |
f12b7f5 | Alexey Radul | 14 February 2011, 13:55:18 UTC | More informative failure reporting (accounts for yet another kind of out-of-resource crash, and for the possibility that the test suite may not have run to completion). | 29 September 2016, 16:14:37 UTC |
ef9a521 | Alexey Radul | 14 February 2011, 13:38:53 UTC | Test supervisor now detects another way in which its subprocess may report dying and marks it as such. | 29 September 2016, 16:14:37 UTC |
f2d5218 | Alexey Radul | 14 February 2011, 13:38:17 UTC | Slightly more pleasant error state reports while the suite is running. | 29 September 2016, 16:14:37 UTC |
5cc08f7 | Alexey Radul | 12 February 2011, 19:34:00 UTC | Noting that double-agent does not have variations. | 29 September 2016, 16:14:37 UTC |
2429e20 | Alexey Radul | 12 February 2011, 19:32:02 UTC | Abstracting the gradient-ascent method in probabilistic-prolog.vlad. | 29 September 2016, 16:14:37 UTC |
24cd59f | Alexey Radul | 12 February 2011, 19:30:56 UTC | Abstracting the derivative-taking strategy in probabilistic-lambda-calculus.vlad. | 29 September 2016, 16:14:37 UTC |
89ea368 | Alexey Radul | 12 February 2011, 19:25:58 UTC | Abstracting the gradient calculations in particle.vlad. | 29 September 2016, 16:14:37 UTC |
7c7e106 | Alexey Radul | 12 February 2011, 19:18:10 UTC | Refactored saddle.vlad to abstract over the different gradient-taking strategies. | 29 September 2016, 16:14:37 UTC |
223340c | Alexey Radul | 12 February 2011, 19:00:41 UTC | The multivariate-argmin in saddle.vlad is an instance of an iterated line search. | 29 September 2016, 16:14:37 UTC |
86b4353 | Alexey Radul | 12 February 2011, 18:57:21 UTC | Making following directions a little friendlier for interpreters (untested, since only hairy things follow directions nowadays). | 29 September 2016, 16:14:37 UTC |