From 4a88a686d5e08404080ed6954b70c0db252aa220 Mon Sep 17 00:00:00 2001
From: Sebastiano Tronto <sebastiano@tronto.net>
Date: Sat, 14 Dec 2024 11:59:22 +0100
Subject: Allow choosing number of threads when calling solve()

---
 src/nissy.c             | 16 +++++++++++++++-
 src/nissy.h             |  4 ++++
 src/solvers/h48/solve.h | 17 ++++++++++-------
 3 files changed, 29 insertions(+), 8 deletions(-)

(limited to 'src')

diff --git a/src/nissy.c b/src/nissy.c
index f933cab..5d009ba 100644
--- a/src/nissy.c
+++ b/src/nissy.c
@@ -489,6 +489,7 @@ nissy_solve(
 	unsigned maxmoves,
 	unsigned maxsols,
 	int optimal,
+	int threads,
 	unsigned long long data_size,
 	const char data[data_size],
 	unsigned sols_size,
@@ -499,6 +500,7 @@ nissy_solve(
 	cube_t c;
 	long long parse_ret;
 	uint8_t h, k;
+	int t;
 
 	if (solver == NULL) {
 		LOG("Error: 'solver' argument is NULL\n");
@@ -522,10 +524,22 @@ nissy_solve(
 		return 0;
 	}
 
+	t = threads == 0 ? THREADS : threads;
+	if (t < 0) {
+		LOG("solve: 'threads' is negative. Please provide a "
+		    "number of threads between 1 and %d\n", THREADS);
+		return NISSY_ERROR_OPTIONS;
+	} 
+	if (t > THREADS) {
+		LOG("solve: 'threads' is above the maximum value of %d\n",
+		    THREADS);
+		return NISSY_ERROR_OPTIONS;
+	}
+
 	if (!strncmp(solver, "h48", 3)) {
 		parse_ret = parse_h48_solver(solver, &h, &k);
 		if (parse_ret == NISSY_OK)
-			return solve_h48(c, minmoves, maxmoves, maxsols,
+			return solve_h48(c, minmoves, maxmoves, maxsols, t,
 			        data_size, data, sols_size, sols, stats);
 		else
 			return parse_ret;
diff --git a/src/nissy.h b/src/nissy.h
index 8394fad..94c604c 100644
--- a/src/nissy.h
+++ b/src/nissy.h
@@ -259,6 +259,9 @@ Parameters:
    maxsols   - The maximum number of solutions.
    optimal   - If set to a non-negative value, the maximum number of moves
                above the optimal solution length.
+   threads   - The number of threads to use. Must be less than or equalt to
+               the value of the compile-time constant THREADS. If set to 0,
+               the default value THREADS will be used.
    data_size - The size of the data buffer.
    data      - The data for the solver. Can be computed with gendata.
    sols_size - The size of the solutions buffer.
@@ -286,6 +289,7 @@ nissy_solve(
 	unsigned maxmoves,
 	unsigned maxsolutions,
 	int optimal,
+	int threads,
 	unsigned long long data_size,
 	const char data[data_size],
 	unsigned sols_size,
diff --git a/src/solvers/h48/solve.h b/src/solvers/h48/solve.h
index 8666007..0969c45 100644
--- a/src/solvers/h48/solve.h
+++ b/src/solvers/h48/solve.h
@@ -38,6 +38,7 @@ typedef struct {
 	int64_t nodes_visited;
 	int64_t table_fallbacks;
 	int64_t table_lookups;
+	int threads;
 	int ntasks;
 	solve_h48_task_t *tasks;
 	int thread_id;
@@ -62,7 +63,7 @@ STATIC int64_t solve_h48_maketasks(
     solve_h48_task_t [static STARTING_CUBES], int *);
 STATIC void *solve_h48_runthread(void *);
 STATIC int64_t solve_h48_dfs(dfsarg_solve_h48_t *);
-STATIC int64_t solve_h48(cube_t, int8_t, int8_t, uint64_t, uint64_t,
+STATIC int64_t solve_h48(cube_t, int8_t, int8_t, uint64_t, int, uint64_t,
     const void *, uint64_t, char *, long long [static NISSY_SIZE_SOLVE_STATS]);
 
 STATIC int64_t
@@ -316,7 +317,7 @@ solve_h48_runthread(void *arg)
 	dfsarg = (dfsarg_solve_h48_t *)arg;
 	cube = dfsarg->start_cube;
 
-	for (i = dfsarg->thread_id; i < dfsarg->ntasks; i += THREADS) {
+	for (i = dfsarg->thread_id; i < dfsarg->ntasks; i += dfsarg->threads) {
 		task = dfsarg->tasks[i];
 		memcpy(dfsarg->moves, task.moves, STARTING_MOVES);
 		dfsarg->cube = cube;
@@ -407,6 +408,7 @@ solve_h48(
 	int8_t minmoves,
 	int8_t maxmoves,
 	uint64_t maxsolutions,
+	int threads,
 	uint64_t data_size,
 	const void *data,
 	uint64_t solutions_size,
@@ -461,7 +463,7 @@ solve_h48(
 	fallback2 = h48data + offset;
 
 	symmask = symmetry_mask(cube);
-	for (i = 0; i < THREADS; i++) {
+	for (i = 0; i < threads; i++) {
 		arg[i] = (dfsarg_solve_h48_t) {
 			.start_cube = cube,
 			.cube = cube,
@@ -481,6 +483,7 @@ solve_h48(
 			.nodes_visited = 0,
 			.table_fallbacks = 0,
 			.table_lookups = 0,
+			.threads = threads,
 			.thread_id = i,
 			.solutions_mutex = &solutions_mutex,
 		};
@@ -505,7 +508,7 @@ solve_h48(
 	if (*arg[0].nsols >= (int64_t)maxsolutions)
 		goto solve_h48_done;
 
-	for (i = 0; i < THREADS; i++) {
+	for (i = 0; i < threads; i++) {
 		arg[i].ntasks = ntasks;
 		arg[i].tasks = tasks;
 	}
@@ -520,12 +523,12 @@ solve_h48(
 		if (d >= 10)
 			LOG("Found %" PRId64 " solutions, searching at depth %"
 			    PRId8 "\n", nsols, d);
-		for (i = 0; i < THREADS; i++) {
+		for (i = 0; i < threads; i++) {
 			arg[i].depth = d;
 			pthread_create(
 			    &thread[i], NULL, solve_h48_runthread, &arg[i]);
 		}
-		for (i = 0; i < THREADS; i++)
+		for (i = 0; i < threads; i++)
 			pthread_join(thread[i], NULL);
 	}
 
@@ -534,7 +537,7 @@ solve_h48_done:
 		goto solve_h48_error_solutions_buffer;
 
 	nodes_visited = table_lookups = table_fallbacks = 0;
-	for (i = 0; i < THREADS; i++) {
+	for (i = 0; i < threads; i++) {
 		nodes_visited += arg[i].nodes_visited;
 		table_fallbacks += arg[i].table_fallbacks;
 		table_lookups += arg[i].table_lookups;
-- 
cgit v1.3