X-Git-Url: https://git.xonotic.org/?p=xonotic%2Fgmqcc.git;a=blobdiff_plain;f=correct.c;h=7e16608a0084c260810a1f9240e757f6d00f0673;hp=bd79c63cdf7ae12892aeb5e9e03e144e2bdc524c;hb=69b55ccc03b56af1f6c05eb45866ab198307487f;hpb=9c8ddb3771ded271e64d710659b615112b42dbcc diff --git a/correct.c b/correct.c index bd79c63..7e16608 100644 --- a/correct.c +++ b/correct.c @@ -2,7 +2,7 @@ * Copyright (C) 2012, 2013 * Dale Weiler * Wolfgang Bumiller - * + * * Permission is hereby granted, free of charge, to any person obtaining a copy of * this software and associated documentation files (the "Software"), to deal in * the Software without restriction, including without limitation the rights to @@ -51,11 +51,11 @@ * Bayesian interpretation. You can read more about it from here: * http://www.celiagreen.com/charlesmccreery/statistics/bayestutorial.pdf * (which is probably the only good online documentation for bayes theroy - * no lie. Everything else just sucks ..) - * + * no lie. Everything else just sucks ..) + * * Bayes' Thereom suggests something like the following: * AC P(I|C) P(C) / P(I) - * + * * However since P(I) is the same for every possibility of I, we can * completley ignore it giving just: * AC P(I|C) P(C) @@ -72,7 +72,7 @@ * 3: AC, the control mechanisim, an enumerator if you will, one that * enumerates all feasible values of C, to determine the one that * gives the greatest probability score. - * + * * In reality the requirement for a more complex expression involving * two seperate models is considerably a waste. But one must recognize * that P(C|I) is already conflating two factors. It's just much simpler @@ -85,8 +85,8 @@ * distance no greater than one. Knowing this we can optimize for most * cases of mistakes without taking a performance hit. Which is what we * base longer edit distances off of. Opposed to the original method of - * I had concieved of checking everything. - * + * I had concieved of checking everything. + * * A little information on additional algorithms used: * * Initially when I implemented this corrector, it was very slow. @@ -106,22 +106,29 @@ * there is also a hash. After 3 million hashes .. you start to get * very slow. To combat this I had suggested burst tries to Blub. * The next day he had implemented them. Sure enough this brought - * down the runtime by a factory > 100% + * down the runtime by a factor > 100% + * + * The trie initially was designed to work on all strings, but later it + * became aparent that not only was this not a requirement. It was also + * slowing down get/sets' for the trie. To fully understand, only + * correct_alpha needs to be understood by the trie system, knowing this + * We can combat the slowness using a very clever but evil optimization. + * By Setting a fixed sized amount of branches for the trie using a + * char-to-index map into the branches. We've complelty made the trie + * accesses entierly constant in lookup time. No really, a lookup is + * literally trie[str[0]] [str[1]] [2] .... .value. + * * * Future Work (If we really need it) * - * Currently we can only distinguishes one source of error in the + * Currently we can only distinguish one source of error in the * language model we use. This could become an issue for identifiers * that have close colliding rates, e.g colate->coat yields collate. * * Currently the error model has been fairly trivial, the smaller the * edit distance the smaller the error. This usually causes some un- * expected problems. e.g reciet->recite yields recipt. For QuakeC - * this could become a problem when lots of identifiers are involved. - * - * Our control mechanisim could use a limit, i.e limit the number of - * sets of edits for distance X. This would also increase execution - * speed considerably. + * this could become a problem when lots of identifiers are involved. */ @@ -168,7 +175,6 @@ static GMQCC_INLINE void correct_pool_delete(void) { static GMQCC_INLINE char *correct_pool_claim(const char *data) { char *claim = util_strdup(data); - correct_pool_delete(); return claim; } @@ -207,7 +213,7 @@ correct_trie_t* correct_trie_new() { return t; } -void correct_trie_del_sub(correct_trie_t *t) { +static GMQCC_INLINE void correct_trie_del_sub(correct_trie_t *t) { size_t i; if (!t->entries) return; @@ -217,7 +223,7 @@ void correct_trie_del_sub(correct_trie_t *t) { mem_d(t->entries); } -void correct_trie_del(correct_trie_t *t) { +static GMQCC_INLINE void correct_trie_del(correct_trie_t *t) { size_t i; if (t->entries) { for (i = 0; i < sizeof(correct_alpha)-1; ++i) @@ -227,7 +233,7 @@ void correct_trie_del(correct_trie_t *t) { mem_d(t); } -void* correct_trie_get(const correct_trie_t *t, const char *key) { +static GMQCC_INLINE void* correct_trie_get(const correct_trie_t *t, const char *key) { const unsigned char *data = (const unsigned char*)key; while (*data) { @@ -239,7 +245,7 @@ void* correct_trie_get(const correct_trie_t *t, const char *key) { return t->value; } -void correct_trie_set(correct_trie_t *t, const char *key, void * const value) { +static GMQCC_INLINE void correct_trie_set(correct_trie_t *t, const char *key, void * const value) { const unsigned char *data = (const unsigned char*)key; while (*data) { if (!t->entries) { @@ -308,9 +314,9 @@ void correct_del(correct_trie_t* dictonary, size_t **data) { * not in the mood to figure out that logic. This is a reminder to * do it, or for someone else to :-) correct_edit however would also * need to take a size_t ** to carry it along (would all the argument - * overhead be worth it?) + * overhead be worth it?) */ -static size_t correct_deletion(const char *ident, char **array) { +static GMQCC_INLINE size_t correct_deletion(const char *ident, char **array) { size_t itr = 0; const size_t len = strlen(ident); @@ -324,7 +330,7 @@ static size_t correct_deletion(const char *ident, char **array) { return itr; } -static size_t correct_transposition(const char *ident, char **array) { +static GMQCC_INLINE size_t correct_transposition(const char *ident, char **array) { size_t itr = 0; const size_t len = strlen(ident); @@ -341,7 +347,7 @@ static size_t correct_transposition(const char *ident, char **array) { return itr; } -static size_t correct_alteration(const char *ident, char **array) { +static GMQCC_INLINE size_t correct_alteration(const char *ident, char **array) { size_t itr = 0; size_t jtr = 0; size_t ktr = 0; @@ -359,7 +365,7 @@ static size_t correct_alteration(const char *ident, char **array) { return ktr; } -static size_t correct_insertion(const char *ident, char **array) { +static GMQCC_INLINE size_t correct_insertion(const char *ident, char **array) { size_t itr = 0; size_t jtr = 0; const size_t len = strlen(ident); @@ -389,11 +395,12 @@ static GMQCC_INLINE size_t correct_size(const char *ident) { return (len) + (len - 1) + (len * (sizeof(correct_alpha)-1)) + ((len + 1) * (sizeof(correct_alpha)-1)); } -static char **correct_edit(const char *ident) { +static GMQCC_INLINE char **correct_edit(const char *ident, size_t **lens) { size_t next; - char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*)); + size_t size = correct_size(ident); + char **find = (char**)correct_pool_alloc(size * sizeof(char*)); - if (!find) + if (!find || !(*lens = (size_t*)correct_pool_alloc(size * sizeof(size_t)))) return NULL; next = correct_deletion (ident, find); @@ -401,44 +408,21 @@ static char **correct_edit(const char *ident) { next += correct_alteration (ident, find+next); /*****/ correct_insertion (ident, find+next); + /* precompute lengths */ + for (next = 0; next < size; next++) + (*lens)[next] = strlen(find[next]); + return find; } -/* - * We could use a hashtable but the space complexity isn't worth it - * since we're only going to determine the "did you mean?" identifier - * on error. - */ -static int correct_exist(char **array, size_t rows, char *ident) { +static GMQCC_INLINE int correct_exist(char **array, register size_t *sizes, size_t rows, char *ident, register size_t len) { size_t itr; - /* - * As an experiment I tried the following assembly for memcmp here: - * - * correct_cmp_loop: - * incl %eax ; eax = LHS - * incl %edx ; edx = LRS - * cmpl %eax, %ebx ; ebx = &LHS[END_POS] - * - * jbe correct_cmp_eq - * movb (%edx), %cl ; micro-optimized even on atoms :-) - * cmpb %cl, (%eax) ; ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - * jg correct_cmp_gt - * jge correct_cmp_loop - * ... - * - * Despite how much optimization went in to this, the speed was the - * being conflicted by the strlen(ident) used for &LHS[END_POS] - * If we could eliminate the strlen with what I suggested on line - * 311 ... we can accelerate this whole damn thing quite a bit. - * - * However there is still something we can do here that does give - * us a little more speed. Although one more branch, we know for - * sure there is at least one byte to compare, if that one byte - * simply isn't the same we can skip the full check. Which means - * we skip a whole strlen call. - */ for (itr = 0; itr < rows; itr++) { - if (!memcmp(array[itr], ident, strlen(ident))) + /* + * We can save tons of calls to memcmp if we simply ignore comparisions + * that we know cannot contain the same length. + */ + if (sizes[itr] == len && !memcmp(array[itr], ident, len)) return 1; } @@ -451,29 +435,38 @@ static GMQCC_INLINE char **correct_known_resize(char **res, size_t *allocated, s if (size < oldallocated) return res; - out = correct_pool_alloc(sizeof(*res) * oldallocated + 32); + out = (char**)correct_pool_alloc(sizeof(*res) * oldallocated + 32); memcpy(out, res, sizeof(*res) * oldallocated); *allocated += 32; return out; } -static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) { +static char **correct_known(correction_t *corr, correct_trie_t* table, char **array, size_t rows, size_t *next) { size_t itr = 0; size_t jtr = 0; size_t len = 0; size_t row = 0; size_t nxt = 8; - char **res = correct_pool_alloc(sizeof(char *) * nxt); - char **end = NULL; + char **res = (char**)correct_pool_alloc(sizeof(char *) * nxt); + char **end = NULL; + size_t *bit = NULL; for (; itr < rows; itr++) { - end = correct_edit(array[itr]); + if (!array[itr][0]) + continue; + if (vec_size(corr->edits) > itr+1) { + end = corr->edits[itr+1]; + bit = corr->lens [itr+1]; + } else { + end = correct_edit(array[itr], &bit); + vec_push(corr->edits, end); + vec_push(corr->lens, bit); + } row = correct_size(array[itr]); - /* removing jtr=0 here speeds it up by 100ms O_o */ for (jtr = 0; jtr < row; jtr++) { - if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) { + if (correct_find(table, end[jtr]) && !correct_exist(res, bit, len, end[jtr], bit[jtr])) { res = correct_known_resize(res, &nxt, len+1); res[len++] = end[jtr]; } @@ -484,7 +477,7 @@ static char **correct_known(correct_trie_t* table, char **array, size_t rows, si return res; } -static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) { +static GMQCC_INLINE char *correct_maximum(correct_trie_t* table, char **array, size_t rows) { char *str = NULL; size_t *itm = NULL; size_t itr = 0; @@ -505,32 +498,50 @@ static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) { * takes a table for the dictonary a vector of sizes (used for internal * probability calculation), and an identifier to "correct". */ -char *correct_str(correct_trie_t* table, const char *ident) { +void correct_init(correction_t *c) +{ + correct_pool_new(); + c->edits = NULL; + c->lens = NULL; +} + +void correct_free(correction_t *c) +{ + vec_free(c->edits); + vec_free(c->lens); + correct_pool_delete(); +} + +char *correct_str(correction_t *corr, correct_trie_t* table, const char *ident) { char **e1 = NULL; char **e2 = NULL; char *e1ident = NULL; char *e2ident = NULL; size_t e1rows = 0; size_t e2rows = 0; - - correct_pool_new(); + size_t *bits = NULL; /* needs to be allocated for free later */ if (correct_find(table, ident)) return correct_pool_claim(ident); if ((e1rows = correct_size(ident))) { - e1 = correct_edit(ident); + if (vec_size(corr->edits) > 0) + e1 = corr->edits[0]; + else { + e1 = correct_edit(ident, &bits); + vec_push(corr->edits, e1); + vec_push(corr->lens, bits); + } if ((e1ident = correct_maximum(table, e1, e1rows))) return correct_pool_claim(e1ident); } - e2 = correct_known(table, e1, e1rows, &e2rows); + e2 = correct_known(corr, table, e1, e1rows, &e2rows); if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows)))) return correct_pool_claim(e2ident); - correct_pool_delete(); return util_strdup(ident); }