ggml: Added OpenMP for multi-threads processing

CMakeLists.txt

@@ -125,6 +125,7 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
+option(LLAMA_OPENMP                          "llama: use OpenMP"                                OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -295,6 +296,17 @@ if (LLAMA_METAL)
         )
 endif()
 
+if (LLAMA_OPENMP)
+    find_package(OpenMP)
+    if(OpenMP_FOUND)
+        message(STATUS "OpenMP found")
+        add_compile_definitions(GGML_USE_OPENMP)
+
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+    endif()
+endif()
+
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)

Makefile

@@ -400,6 +400,12 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif # LLAMA_NO_ACCELERATE
 
+ifdef LLAMA_OPENMP
+	MK_CPPFLAGS += -DGGML_USE_OPENMP
+	MK_CFLAGS   += -fopenmp
+	MK_CXXFLAGS += -fopenmp
+endif # LLAMA_OPENMP
+
 ifdef LLAMA_OPENBLAS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)

ggml.c

@@ -28,6 +28,10 @@
 #include <syscall.h>
 #endif
 
+#ifdef GGML_USE_OPENMP
+#include <omp.h>
+#endif
+
 #ifdef GGML_USE_METAL
 #include <unistd.h>
 #endif
@@ -19661,6 +19665,48 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
     return cplan;
 }
 
+static enum ggml_status ggml_graph_compute_parallel(struct ggml_compute_state * workers, int n_threads){
+    enum ggml_status compute_status = GGML_STATUS_SUCCESS;
+
+#ifdef GGML_USE_OPENMP
+#pragma omp parallel num_threads(n_threads)
+    {
+        ggml_graph_compute_thread(&workers[omp_get_thread_num()]);
+    }
+#else
+    // create thread pool
+    if (n_threads > 1) {
+        for (int j = 1; j < n_threads; ++j) {
+            const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
+            GGML_ASSERT(rc == 0);
+            UNUSED(rc);
+        }
+    }
+
+    // this is a work thread too
+    ggml_graph_compute_thread(&workers[0]);
+
+    // don't leave affinity set on the main thread
+    clear_numa_thread_affinity();
+
+    // join or kill thread pool
+    if (n_threads > 1) {
+        for (int j = 1; j < n_threads; j++) {
+            const int rc = ggml_thread_join(workers[j].thrd, NULL);
+            GGML_ASSERT(rc == 0);
+            UNUSED(rc);
+        }
+    }
+#endif
+    for (int j = 0; j < n_threads; j++) {
+        if (workers[j].ec != GGML_STATUS_SUCCESS) {
+            compute_status = workers[j].ec;
+            break;
+        }
+    }
+    return compute_status;
+}
+
 enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
     {
         GGML_ASSERT(cplan);
@@ -19687,47 +19733,20 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         /*.current_chunk;          =*/ 0,
     };
     struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads);
-
-    // create thread pool
-    if (n_threads > 1) {
-        for (int j = 1; j < n_threads; ++j) {
-            workers[j] = (struct ggml_compute_state) {
-                .thrd   = 0,
-                .ith = j,
-                .shared = &state_shared,
-                .ec = GGML_STATUS_SUCCESS,
-            };
-
-            const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
-            GGML_ASSERT(rc == 0);
-            UNUSED(rc);
-        }
-    }
-
-    workers[0].ith = 0;
-    workers[0].shared = &state_shared;
-    workers[0].ec = GGML_STATUS_SUCCESS;
-
     const int64_t perf_start_cycles  = ggml_perf_cycles();
     const int64_t perf_start_time_us = ggml_perf_time_us();
 
-    // this is a work thread too
+    for (int j = 0; j < n_threads; ++j) {
-    ggml_graph_compute_thread(&workers[0]);
+        workers[j] = (struct ggml_compute_state) {
-    enum ggml_status compute_status = workers[0].ec;
+            .thrd   = 0,
-
+            .ith    = j,
-    // don't leave affinity set on the main thread
+            .shared = &state_shared,
-    clear_numa_thread_affinity();
+            .ec     = GGML_STATUS_SUCCESS,
-
+        };
-    // join or kill thread pool
-    if (n_threads > 1) {
-        for (int j = 1; j < n_threads; j++) {
-            const int rc = ggml_thread_join(workers[j].thrd, NULL);
-            GGML_ASSERT(rc == 0);
-            if (workers[j].ec != GGML_STATUS_SUCCESS)
-                compute_status = workers[j].ec;
-        }
     }
 
+    enum ggml_status compute_status = ggml_graph_compute_parallel(workers, n_threads);
+
     // performance stats (graph)
     {
         int64_t perf_cycles_cur  = ggml_perf_cycles()  - perf_start_cycles;