threadpool: reduce pause/resume/wakeup overhead in common cases

We now start threadpool in paused state only if we have two.
The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead.

Author: max-krasnyansky

common/common.cpp

@@ -2602,6 +2602,7 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     tpp.prio       = params.priority;
     tpp.poll       = params.poll;
     tpp.strict_cpu = params.strict_cpu;
+    tpp.paused     = false;
 
     return tpp;
 }

examples/main/main.cpp

@@ -230,17 +230,6 @@ int main(int argc, char ** argv) {
     struct ggml_threadpool_params tpp =
             ggml_threadpool_params_from_cpu_params(params.cpuparams);
 
-    struct ggml_compute_threadpool * threadpool = ggml_create_threadpool(&tpp);
-    if (!threadpool) {
-        LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
-        exit(1);
-    }
-
-    llama_attach_threadpool(ctx, threadpool);
-    if (ctx_guidance) {
-        llama_attach_threadpool(ctx_guidance, threadpool);
-    }
-
     struct ggml_compute_threadpool * threadpool_batch = NULL;
     if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
         threadpool_batch = ggml_create_threadpool(&tpp_batch);
@@ -253,6 +242,20 @@ int main(int argc, char ** argv) {
         if (ctx_guidance) {
             llama_attach_batch_threadpool(ctx_guidance, threadpool_batch);
         }
+
+        // Start the non-batch threadpool in the paused state
+        tpp.paused = true;
+    }
+
+    struct ggml_compute_threadpool * threadpool = ggml_create_threadpool(&tpp);
+    if (!threadpool) {
+        LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+        exit(1);
+    }
+
+    llama_attach_threadpool(ctx, threadpool);
+    if (ctx_guidance) {
+        llama_attach_threadpool(ctx_guidance, threadpool);
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);

ggml/include/ggml.h

@@ -633,6 +633,7 @@ extern "C" {
         int32_t prio;
         bool    poll;
         bool    strict_cpu;
+        bool    paused;
     };
 
     struct ggml_compute_threadpool;     // forward declaration, see ggml.c

ggml/src/ggml.c

@@ -18872,14 +18872,27 @@ void ggml_release_threadpool(struct ggml_compute_threadpool* threadpool) {
     GGML_ALIGNED_FREE(threadpool);
 }
 
+#ifndef GGML_USE_OPENMP
+// pause/resume must be called under mutex
+static void __ggml_pause_threadpool(struct ggml_compute_threadpool * threadpool) {
+    GGML_PRINT_DEBUG("Pausing threadpool\n");
+    threadpool->pause = true;
+    ggml_cond_broadcast(&threadpool->cond);
+}
+
+static void __ggml_resume_threadpool(struct ggml_compute_threadpool * threadpool) {
+    GGML_PRINT_DEBUG("Resuming threadpool\n");
+    threadpool->pause = false;
+    ggml_cond_broadcast(&threadpool->cond);
+}
+#endif
+
 void ggml_pause_threadpool(struct ggml_compute_threadpool * threadpool) {
 #ifndef GGML_USE_OPENMP
     GGML_ASSERT(!threadpool->disposable);
-    GGML_PRINT_DEBUG("Pausing threadpool\n");
     ggml_mutex_lock(&threadpool->mutex);
     if (!threadpool->pause) {
-        threadpool->pause = true;
-        ggml_cond_broadcast(&threadpool->cond);
+       __ggml_pause_threadpool(threadpool);
     }
     ggml_mutex_unlock(&threadpool->mutex);
 #else
@@ -18890,12 +18903,9 @@ void ggml_pause_threadpool(struct ggml_compute_threadpool * threadpool) {
 void ggml_resume_threadpool(struct ggml_compute_threadpool * threadpool) {
 #ifndef GGML_USE_OPENMP
     GGML_ASSERT(!threadpool->disposable);
-    GGML_PRINT_DEBUG("Resuming threadpool\n");
-
     ggml_mutex_lock(&threadpool->mutex);
     if (threadpool->pause) {
-        threadpool->pause = false;
-        ggml_cond_broadcast(&threadpool->cond);
+       __ggml_resume_threadpool(threadpool);
     }
     ggml_mutex_unlock(&threadpool->mutex);
 #else
@@ -19237,7 +19247,7 @@ static struct ggml_compute_threadpool * ggml_create_threadpool_impl(
         threadpool->n_barrier_passed = 0;
         threadpool->current_chunk    = 0;
         threadpool->stop             = false;
-        threadpool->pause            = disposable ? false : true;
+        threadpool->pause            = disposable ? false : tpp->paused;
         threadpool->new_work         = false;
         threadpool->workers          = NULL;
         threadpool->n_threads_max    = tpp->n_threads;
@@ -19327,9 +19337,10 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         struct ggml_threadpool_params ttp = {
             .mask_specified = false,
             .n_threads      = n_threads,
-            .prio           = 1,
+            .prio           = 0,
             .poll           = false,
-            .strict_cpu     = false
+            .strict_cpu     = false,
+            .paused         = false
         };
 
         threadpool = ggml_create_threadpool_impl(&ttp, true, cgraph, cplan);
@@ -19383,10 +19394,19 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         if (!threadpool->poll) {
             ggml_mutex_lock(&threadpool->mutex);
             threadpool->new_work = true;
-            ggml_cond_broadcast(&threadpool->cond);
+            if (threadpool->pause) {
+               __ggml_resume_threadpool(threadpool);
+            } else {
+               ggml_cond_broadcast(&threadpool->cond);
+            }
             ggml_mutex_unlock(&threadpool->mutex);
         } else {
             threadpool->new_work = true;
+            if (threadpool->pause) {
+                ggml_mutex_lock(&threadpool->mutex);
+                __ggml_resume_threadpool(threadpool);
+                ggml_mutex_unlock(&threadpool->mutex);
+            }
         }
     }
     // this is a work thread too

src/llama.cpp

@@ -15539,17 +15539,14 @@ static std::pair<int32_t, ggml_compute_threadpool_t> llama_swap_threadpools(
         // Switch between the 2 threadpools as needed
         if (n_tokens > 1) {
             ggml_pause_threadpool(lctx.threadpool);
-            ggml_resume_threadpool(lctx.threadpool_batch);
             threadpool = lctx.threadpool_batch;
             n_threads = cparams.n_threads_batch;
         } else {
             ggml_pause_threadpool(lctx.threadpool_batch);
-            ggml_resume_threadpool(lctx.threadpool);
             threadpool = lctx.threadpool;
             n_threads = cparams.n_threads;
         }
     } else if (lctx.threadpool) {
-        ggml_resume_threadpool(lctx.threadpool);
         threadpool = lctx.threadpool;
         n_threads = cparams.n_threads;
     }