cont : reuse current_chunk from ggml_threadpool

ggml-ci

Author: ggerganov

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -503,6 +503,9 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * threadpool, int value);
+int  ggml_threadpool_chunk_add(struct ggml_threadpool * threadpool, int value);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/ggml-cpu.c

@@ -559,6 +559,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * threadpool, int value) {
+    atomic_store_explicit(&threadpool->current_chunk, value, memory_order_release);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * threadpool, int value) {
+    return atomic_fetch_add_explicit(&threadpool->current_chunk, value, memory_order_acq_rel);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -406,6 +406,8 @@ class tinyBLAS {
 
         if (params->ith == 0) {
             GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
         }
 
         ggml_barrier(params->threadpool);
@@ -434,7 +436,7 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            job += params->nth;
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
 
         ggml_barrier(params->threadpool);