threadpool: skip polling for unused threads

Currently all threads do N polling rounds even if only 1 thread is active (n_threads_cur == 1).
This commit adds a check to skip the polling for unused threads (ith >= n_threads_cur).

n_threads_cur is now an atomic_int to explicitly tell thread sanitizer that it is written
from one thread and read from other threads (not a race conditions).

Author: max-krasnyansky

ggml/src/ggml.c

@@ -2015,7 +2015,7 @@ struct ggml_threadpool {
 
     struct ggml_compute_state * workers;   // per thread state
     int          n_threads_max; // number of threads in the pool
-    int          n_threads_cur; // number of threads used in the current graph
+    atomic_int   n_threads_cur; // number of threads used in the current graph
 
     int32_t      prio;        // Scheduling priority
     uint32_t     poll;        // Polling level (0 - no polling)
@@ -19967,15 +19967,20 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
 #ifndef GGML_USE_OPENMP
 
-static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
+static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    return (state->ith < threadpool->n_threads_cur);
+}
+
+static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (state->pending || threadpool->stop || threadpool->pause) { return true; }
 
     // check for new graph/work
     int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
     if (new_graph != state->last_graph) {
-        state->pending    = (state->ith < threadpool->n_threads_cur);
+        state->pending    = ggml_graph_compute_thread_active(state);
         state->last_graph = new_graph;
     }
 
@@ -19985,11 +19990,16 @@ static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
 static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
+    // Skip polling for unused threads
+    if (!ggml_graph_compute_thread_active(state)) {
+        return state->pending;
+    }
+
     // This seems to make 0 ... 100 a decent range for polling level across modern processors.
     // Perhaps, we can adjust it dynamically based on load and things.
     const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;
 
-    for (uint64_t i=0; !ggml_graph_compute_ready(state) && i<n_rounds; i++) {
+    for (uint64_t i=0; !ggml_graph_compute_thread_ready(state) && i < n_rounds; i++) {
         // No new work. Keep polling.
         ggml_thread_cpu_relax();
     }
@@ -20005,9 +20015,9 @@ static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state *
     }
 
     ggml_mutex_lock_shared(&threadpool->mutex);
-    while (!ggml_graph_compute_ready(state)) {
+    while (!ggml_graph_compute_thread_ready(state)) {
         // No new work. Wait for the signal.
-        GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith);
+        GGML_PRINT_DEBUG("thread #%d waiting for work (sleeping)\n", state->ith);
         ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
     }
     ggml_mutex_unlock_shared(&threadpool->mutex);
@@ -20054,12 +20064,17 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
 }
 
 // Start processing new graph
-static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool)
+static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int n_threads)
 {
     // always take the mutex here because the worker threads are doing hybrid poll/wait
 
     ggml_mutex_lock(&threadpool->mutex);
 
+    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+
+    // Update the number of active threads
+    threadpool->n_threads_cur = n_threads;
+
     atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
 
     if (threadpool->pause) {
@@ -20194,15 +20209,10 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         // No worker threads should be accessing the parameters below at this stage
         threadpool->cgraph           = cgraph;
         threadpool->cplan            = cplan;
-        threadpool->n_threads_cur    = n_threads;
         threadpool->current_chunk    = 0;
         threadpool->ec               = GGML_STATUS_SUCCESS;
     }
 
-    if (n_threads > threadpool->n_threads_max) {
-        GGML_PRINT("WARNING: cplan is requesting more threads than the threadpool contains. Expect a bad time!\n");
-    }
-
 #ifdef GGML_USE_OPENMP
     if (n_threads > 1) {
         #pragma omp parallel num_threads(n_threads)
@@ -20220,8 +20230,13 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         ggml_graph_compute_thread(&threadpool->workers[0]);
     }
 #else
+    if (n_threads > threadpool->n_threads_max) {
+        GGML_PRINT("WARNING: cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
+        n_threads = threadpool->n_threads_max;
+    }
+
     // Kick all threads to start the new graph
-    ggml_graph_compute_kickoff(threadpool);
+    ggml_graph_compute_kickoff(threadpool, n_threads);
 
     // This is a work thread too
     ggml_graph_compute_thread(&threadpool->workers[0]);