CFL reachability problem using Kronecker product by Mamontyonok · Pull Request #10 · SparseLinearAlgebra/LAGraph

Mamontyonok · 2026-03-06T17:13:50Z

The basic version of the CFLR algorithm is based on the Kronecker product

gsvgit

Add more explanatory comments. For example look at this request

gsvgit · 2026-03-10T07:20:22Z

+    GrB_Index nvals_old = 0, nvals_new = 0;
+    GrB_Matrix_nvals(&nvals_old, A);
+    while (true) {
+        GrB_mxm(A, NULL, GrB_LOR, GrB_LOR_LAND_SEMIRING_BOOL, A, A, NULL);


Why not ANY_PAIR?

gsvgit · 2026-03-10T07:21:13Z

+    char *msg
+)
+{
+    GrB_Matrix M3 = NULL;


Naming. What is M3?

gsvgit · 2026-03-10T07:26:36Z

+    GrB_Index r_dim = rsm->state_count;
+    GrB_Index kronecker_dim = r_dim * g_dim;
+
+    for (int64_t nt = 0; nt < rsm->nonterminal_count; ++nt) {


It is hard to realize this loop. We iterate through all non-terminals. Each non-terminal must has related box and this box must has at least one final state. So, it is unclear what exactly we do in this loop.

Am I right, that you just handle boxes where the same state is both final and start?

yes, that's it

gsvgit · 2026-03-10T07:28:35Z

+        GrB_Matrix_clear(M3);
+
+        for (int64_t t = 0; t < terms_count; ++t) {
+            GrB_kronecker(M3, NULL, GrB_LOR, GrB_LAND,


Why not ANY_PAIR?

gsvgit · 2026-03-10T07:37:26Z

+            GrB_Matrix_extractTuples_BOOL(i, j, V, &m3_nvals, M3);
+
+            for (GrB_Index k = 0; k < m3_nvals; ++k) {
+                GrB_Index s = i[k] / g_dim; 


gsvgit · 2026-03-10T07:39:05Z

+                GrB_Index x = i[k] % g_dim; 
+                GrB_Index y = j[k] % g_dim; 
+
+                for (GrB_Index nt = 0; nt < rsm->nonterminal_count; ++nt) {


Can it be done using matrix operations like select or reduce?

gsvgit · 2026-04-07T10:30:15Z

+    GrB_Matrix_new(&BaseGraph, GrB_BOOL, kronecker_dim, kronecker_dim);
+    GrB_Matrix_new(&DeltaM, GrB_BOOL, kronecker_dim, kronecker_dim);
+
+    for (int64_t t = 0; t < terms_count; ++t) {


Забыли почистить код? Что-то тут странное, начинач со строки 91

Да, видимо когда менял код, продублировал этот цикл

gsvgit · 2026-04-07T10:57:50Z

+
+        GrB_Matrix_dup(&CombinedGraph, M_graph);
+
+        transitive_closure_inplace(CombinedGraph);


Если я правильн понимаю, что здесь происходит, то можно иначе пересчитывать транзитивыное замыкание. Пусть TC -- результат замыкания с предыдущей итерации, D -- дельта на этой итерации. Тогда можно посчиать transitive_closure_inplace(TC + D). Должно чуть меньше итераций требовать. Но надо проверять, что будет быстрее.

homka122 · 2026-04-25T16:44:12Z

+    for (int64_t nt = 0; nt < rsm->nonterminal_count; ++nt) {
+        GrB_Matrix_new(&outputs[nt], GrB_BOOL, g_dim, g_dim);
+        GrB_Matrix_new(&delta_outputs[nt], GrB_BOOL, g_dim, g_dim);
+
+        GrB_Index s = rsm->start_states[nt];
+        bool is_final = false;
+        GrB_Vector_extractElement_BOOL(&is_final, rsm->final_states[nt], s);
+
+        if (is_final) {
+            for (GrB_Index v = 0; v < g_dim; ++v) {
+                GrB_Matrix_setElement_BOOL(outputs[nt], true, v, v);
+                GrB_Matrix_setElement_BOOL(delta_outputs[nt], true, v, v);
+            }
+        }
+    }


Hi, I wanted to add your algo to CFG_Bench, and that piece of code took a lot of time: around 3 seconds to process 1k iterations after 55k/256k iterations had been processed. The processing time increases linearly with the current iteration count.

I think it just adds the identity matrix to the outputs[nt] and delta_outputs[nt] matrices, so this code resolves the task and works in under 1 second on the c_alias init.g graph:

GrB_Vector v_diag; GrB_Vector_new(&v_diag, GrB_BOOL, g_dim); GrB_Vector_assign_BOOL(v_diag, GrB_NULL, GrB_NULL, true, GrB_ALL, g_dim, NULL); for (int64_t nt = 0; nt < rsm->nonterminal_count; ++nt) { GrB_Matrix_new(&outputs[nt], GrB_BOOL, g_dim, g_dim); GrB_Matrix_new(&delta_outputs[nt], GrB_BOOL, g_dim, g_dim); GrB_Index s = rsm->start_states[nt]; bool is_final = false; GrB_Vector_extractElement_BOOL(&is_final, rsm->final_states[nt], s); if (is_final) { GrB_Matrix_diag(&outputs[nt], v_diag, 0); GrB_Matrix_diag(&delta_outputs[nt], v_diag, 0); } } GrB_free(&v_diag);

Mamontyonok added 2 commits March 6, 2026 19:22

init_commit_reachability

4c90003

Update LAGraph_CFL_Kron.c

4191238

gsvgit reviewed Mar 10, 2026

View reviewed changes

Mamontyonok added 5 commits March 11, 2026 17:59

LOR_LAND -> ANY_PAIR

043d17c

the_idea_of_extracting_blocks

c1c32f8

naming

f41e334

using_the_GrB_LOR_accumulator_replaces_the_need_for_GrB_eWiseAdd

efc0df4

The_incremental_version_of_the_algorithm

42e490f

gsvgit reviewed Apr 7, 2026

View reviewed changes

code_duplication_patch

08de903

homka122 reviewed Apr 25, 2026

View reviewed changes


		GrB_Matrix_dup(&CombinedGraph, M_graph);

		transitive_closure_inplace(CombinedGraph);

Conversation

Mamontyonok commented Mar 6, 2026

Uh oh!

gsvgit left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

homka122 Apr 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

homka122 Apr 25, 2026 •

edited

Loading