summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--.github/workflows/build.yml10
-rw-r--r--CMakeLists.txt12
-rw-r--r--Makefile8
-rw-r--r--ggml.c111
4 files changed, 103 insertions, 38 deletions
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 7b616281..e824136a 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -294,12 +294,22 @@ jobs:
- name: Build
id: cmake_build
+ if: ${{ matrix.sanitizer != 'THREAD' }}
run: |
mkdir build
cd build
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+ - name: Build (no OpenMP)
+ id: cmake_build_no_openmp
+ if: ${{ matrix.sanitizer == 'THREAD' }}
+ run: |
+ mkdir build
+ cd build
+ cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_OPENMP=OFF
+ cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+
- name: Test
id: cmake_test
run: |
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0f07f9a5..620305ca 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -126,6 +126,7 @@ 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" ON)
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")
@@ -296,6 +297,17 @@ if (LLAMA_METAL)
)
endif()
+if (LLAMA_OPENMP)
+ find_package(OpenMP)
+ if (OpenMP_FOUND)
+ message(STATUS "OpenMP found")
+ add_compile_definitions(GGML_USE_OPENMP)
+ set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+ else()
+ message(WARNING "OpenMP not found")
+ endif()
+endif()
+
if (LLAMA_BLAS)
if (LLAMA_STATIC)
set(BLA_STATIC ON)
diff --git a/Makefile b/Makefile
index dca2aafd..43b94db0 100644
--- a/Makefile
+++ b/Makefile
@@ -57,6 +57,8 @@ ifeq ($(UNAME_S),Darwin)
LLAMA_METAL := 1
endif
+ LLAMA_NO_OPENMP := 1
+
ifneq ($(UNAME_P),arm)
SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
ifeq ($(SYSCTL_M),1)
@@ -405,6 +407,12 @@ ifndef LLAMA_NO_ACCELERATE
endif
endif # LLAMA_NO_ACCELERATE
+ifndef LLAMA_NO_OPENMP
+ MK_CPPFLAGS += -DGGML_USE_OPENMP
+ MK_CFLAGS += -fopenmp
+ MK_CXXFLAGS += -fopenmp
+endif # LLAMA_NO_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)
diff --git a/ggml.c b/ggml.c
index f479dc3e..000d9db7 100644
--- a/ggml.c
+++ b/ggml.c
@@ -5,6 +5,7 @@
#include "ggml-quants.h"
#include "ggml.h"
+
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
@@ -28,6 +29,10 @@
#include <syscall.h>
#endif
+#ifdef GGML_USE_OPENMP
+#include <omp.h>
+#endif
+
#ifdef GGML_USE_METAL
#include <unistd.h>
#endif
@@ -1756,7 +1761,7 @@ struct ggml_compute_state_shared {
int64_t perf_node_start_cycles;
int64_t perf_node_start_time_us;
- const int n_threads;
+ int n_threads;
// synchronization primitives
atomic_int n_active; // num active threads
@@ -19670,6 +19675,59 @@ 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
+ if (n_threads > 1) {
+ #pragma omp parallel num_threads(n_threads)
+ {
+ #pragma omp single
+ {
+ // update the number of threads from the actual number of threads that we got from OpenMP
+ n_threads = omp_get_num_threads();
+ workers[0].shared->n_threads = n_threads;
+ workers[0].shared->n_active = n_threads;
+ }
+ ggml_graph_compute_thread(&workers[omp_get_thread_num()]);
+ }
+ } else {
+ ggml_graph_compute_thread(&workers[0]);
+ }
+#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]);
+
+ // 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
+ // don't leave affinity set on the main thread
+ clear_numa_thread_affinity();
+
+ 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);
@@ -19680,7 +19738,11 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
}
}
- const int n_threads = cplan->n_threads;
+ int n_threads = cplan->n_threads;
+
+#if defined(GGML_USE_OPENMP)
+ n_threads = MIN(n_threads, omp_get_max_threads());
+#endif
struct ggml_compute_state_shared state_shared = {
/*.cgraph =*/ cgraph,
@@ -19696,47 +19758,20 @@ 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
- ggml_graph_compute_thread(&workers[0]);
- enum ggml_status compute_status = workers[0].ec;
-
- // 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);
- if (workers[j].ec != GGML_STATUS_SUCCESS)
- compute_status = workers[j].ec;
- }
+ for (int j = 0; j < n_threads; ++j) {
+ workers[j] = (struct ggml_compute_state) {
+ .thrd = 0,
+ .ith = j,
+ .shared = &state_shared,
+ .ec = GGML_STATUS_SUCCESS,
+ };
}
+ 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;
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-17 21:53:37 +0000