Revert "threadpool : skip polling for unused threads (ggml-org#9461)"

This reverts commit 0226613.

Author: Nexesenex

ggml/src/ggml.c

@@ -2014,11 +2014,10 @@ struct ggml_threadpool {
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
     atomic_bool pause;        // Used for pausing the threadpool or individual threads
-    atomic_bool abort;        // Used for aborting processing of a graph
 
     struct ggml_compute_state * workers;   // per thread state
     int          n_threads_max; // number of threads in the pool
-    atomic_int   n_threads_cur; // number of threads used in the current graph
+    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)
@@ -3182,36 +3181,41 @@ inline static void ggml_critical_section_start(void) {
     }
 }
 
-static void ggml_barrier(struct ggml_threadpool * tp) {
-    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
-    if (n_threads == 1) {
+#ifdef GGML_USE_OPENMP
+static void ggml_barrier(struct ggml_threadpool * threadpool) {
+    if (threadpool->n_threads_cur == 1) {
         return;
     }
 
-#ifdef GGML_USE_OPENMP
     #pragma omp barrier
+}
 #else
-    int n_passed = atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed);
+static void ggml_barrier(struct ggml_threadpool * threadpool) {
+    if (threadpool->n_threads_cur == 1) {
+        return;
+    }
+
+    atomic_int * n_barrier = &threadpool->n_barrier;
+    atomic_int * n_barrier_passed = &threadpool->n_barrier_passed;
 
-    // enter barrier (full seq-cst fence)
-    int n_barrier = atomic_fetch_add_explicit(&tp->n_barrier, 1, memory_order_seq_cst);
+    int n_threads = threadpool->n_threads_cur;
+    int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed);
 
-    int last = 0;
-    if (n_barrier == (n_threads - 1)) {
+    if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) {
         // last thread
-        atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
-        last = 1;
+        atomic_store(n_barrier, 0);
+        atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed);
     } else {
         // wait for other threads
-        while (atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed) == n_passed) {
+        while (true) {
+            if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) {
+                return;
+            }
             ggml_thread_cpu_relax();
         }
     }
-
-    // exit barrier (full seq-cst fence)
-    atomic_fetch_add_explicit(&tp->n_barrier_passed, last, memory_order_seq_cst);
-#endif
 }
+#endif
 
 // TODO: make this somehow automatically executed
 //       some sort of "sentry" mechanism
@@ -20181,84 +20185,64 @@ struct ggml_cplan ggml_graph_plan(
 
 static thread_ret_t ggml_graph_compute_thread(void * data) {
     struct ggml_compute_state * state = (struct ggml_compute_state *) data;
-    struct ggml_threadpool    * tp    = state->threadpool;
 
-    const struct ggml_cgraph * cgraph = tp->cgraph;
-    const struct ggml_cplan  * cplan  = tp->cplan;
+    const struct ggml_cgraph * cgraph = state->threadpool->cgraph;
+    const struct ggml_cplan  * cplan  = state->threadpool->cplan;
 
     set_numa_thread_affinity(state->ith);
 
     struct ggml_compute_params params = {
         /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
+        /*.nth       =*/ state->threadpool->n_threads_cur,
         /*.wsize     =*/ cplan->work_size,
         /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ tp,
+        /*.threadpool=*/ state->threadpool,
     };
 
-    for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
+    for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
         ggml_compute_forward(&params, node);
 
-        if (state->ith == 0 && cplan->abort_callback &&
-                cplan->abort_callback(cplan->abort_callback_data)) {
-            tp->abort = true;
-            tp->ec    = GGML_STATUS_ABORTED;
+        if (state->ith == 0 && cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
+            state->threadpool->ec = GGML_STATUS_ABORTED;
         }
 
         ggml_barrier(state->threadpool);
+
+        if (state->threadpool->ec != GGML_STATUS_SUCCESS) {
+            break;
+        }
     }
 
     return 0;
 }
 
 #ifndef GGML_USE_OPENMP
 
-// check if thread is active
-static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
-    struct ggml_threadpool * threadpool = state->threadpool;
-    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
-    return (state->ith < n_threads);
-}
-
-// check if thread is ready to proceed (exit from polling or sleeping)
-static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
+static inline bool ggml_graph_compute_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    = ggml_graph_compute_thread_active(state);
+        state->pending    = (state->ith < threadpool->n_threads_cur);
         state->last_graph = new_graph;
     }
 
     return state->pending;
 }
 
-// sync thread state after polling
-static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * state) {
-    struct ggml_threadpool * threadpool = state->threadpool;
-    // this should just be atomic_thread_fence(seq_cst) but it confuses thread-sanitizer
-    // so instead we just use a dummy read-modify-write
-    atomic_fetch_add_explicit(&threadpool->n_graph, 0, memory_order_seq_cst);
-}
-
 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_thread_ready(state) && i < n_rounds; i++) {
+    for (uint64_t i=0; !ggml_graph_compute_ready(state) && i<n_rounds; i++) {
         // No new work. Keep polling.
         ggml_thread_cpu_relax();
     }
@@ -20270,14 +20254,13 @@ static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state *
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (ggml_graph_compute_poll_for_work(state)) {
-        ggml_graph_compute_thread_sync(state);
         return state->pending;
     }
 
     ggml_mutex_lock_shared(&threadpool->mutex);
-    while (!ggml_graph_compute_thread_ready(state)) {
+    while (!ggml_graph_compute_ready(state)) {
         // No new work. Wait for the signal.
-        GGML_PRINT_DEBUG("thread #%d waiting for work (sleeping)\n", state->ith);
+        GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith);
         ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
     }
     ggml_mutex_unlock_shared(&threadpool->mutex);
@@ -20324,20 +20307,13 @@ 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, int n_threads)
+static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool)
 {
-    // Always take the mutex here because the worker threads are doing hybrid poll/wait
+    // 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
-    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
-
-    // Indicate the graph is ready to be processed
-    // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
-    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);
+    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
 
     if (threadpool->pause) {
        // Update main thread prio and affinity to match the threadpool settings
@@ -20396,7 +20372,6 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
         threadpool->current_chunk    = 0;
         threadpool->stop             = false;
         threadpool->pause            = tpp->paused;
-        threadpool->abort            = false;
         threadpool->workers          = NULL;
         threadpool->n_threads_max    = tpp->n_threads;
         threadpool->n_threads_cur    = tpp->n_threads;
@@ -20472,11 +20447,15 @@ 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->abort            = false;
         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)
@@ -20485,7 +20464,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             {
                 // update the number of threads from the actual number of threads that we got from OpenMP
                 n_threads = omp_get_num_threads();
-                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+                threadpool->n_threads_cur = n_threads;
             }
 
             ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
@@ -20495,13 +20474,8 @@ 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, n_threads);
+    ggml_graph_compute_kickoff(threadpool);
 
     // This is a work thread too
     ggml_graph_compute_thread(&threadpool->workers[0]);

tests/CMakeLists.txt

@@ -119,7 +119,6 @@ llama_target_and_test(test-grammar-parser.cpp)
 llama_target_and_test(test-llama-grammar.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
 llama_target_and_test(test-grad0.cpp)
-llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)