try fix Windows CMake build errors;

try fix ubuntu build error: undefined reference to 'sqrtf';
try fix LeakSanitizer error;

Author: mqy

CMakeLists.txt

@@ -158,7 +158,7 @@ if (LLAMA_BLAS)
     endif()
     set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
-    if (BLAS_FOUND) 
+    if (BLAS_FOUND)
         message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
         add_compile_options(${BLAS_LINKER_FLAGS})

Makefile

@@ -290,8 +290,8 @@ ggml-tune.o: ggml-tune.c ggml.h ggml-tune.h
 mulmat-tune: examples/mulmat-tune/mulmat-tune.c ggml.o $(OBJS)
 	$(CC)  $(CFLAGS) $^ -o mulmat-tune $(LDFLAGS)
 
-test-ggml-tune: tests/test-ggml-tune.c ggml.o $(OBJS)
-	$(CC)  $(CFLAGS) $^ -o tests/test-ggml-tune $(LDFLAGS)
+test-mulmat-tune: tests/test-mulmat-tune.c ggml.o $(OBJS)
+	$(CC)  $(CFLAGS) $^ -o tests/test-mulmat-tune $(LDFLAGS)
 
 .PHONY: tests clean
 tests:

examples/mulmat-tune/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET mulmat-tune)
 add_executable(${TARGET} mulmat-tune.c)
-target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml m ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE c_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/mulmat-tune/mulmat-tune.c

@@ -301,6 +301,8 @@ int main(int argc, char **argv) {
             exit(1);
         }
 
+        ggml_mulmat_tune_free(&tune);
+
         if (arg_file != NULL) {
             printf("result was written to %s\n", arg_file);
         }

ggml-tune.c

@@ -87,9 +87,24 @@ int ggml_mulmat_tune_setup_model(struct ggml_mulmat_tune *tune,
     return 0;
 }
 
+int ggml_mulmat_tune_free(struct ggml_mulmat_tune *tune) {
+    for (int i = 0; i < tune->n_shapes; i++) {
+        struct ggml_mulmat_tune_shape *shape = &tune->shapes[i];
+        GGML_ASSERT(shape);
+        GGML_ASSERT(shape->arr_m);
+        GGML_ASSERT(shape->items);
+        free(shape->arr_m);
+        free(shape->items);
+    }
+}
+
 int ggml_mulmat_tune_validate(struct ggml_mulmat_tune *tune, const char *model,
                               int type) {
     enum ggml_backend backend = ggml_auto_detect_backend();
+
+    GGML_ASSERT(backend > GGML_BACKEND_CPU);
+    GGML_ASSERT(tune->backend_vendor);
+
     const char *backend_vendor = ggml_get_backend_vendor();
 
     int rc = 0;
@@ -100,7 +115,7 @@ int ggml_mulmat_tune_validate(struct ggml_mulmat_tune *tune, const char *model,
         rc = -2;
     } else if ((int)backend != tune->backend) {
         rc = -3;
-    } else if (strcmp(backend_vendor, tune->backend_vendor) != 0) {
+    } else if (backend_vendor == NULL || strcmp(backend_vendor, tune->backend_vendor) != 0) {
         rc = -4;
     } else {
         // TODO

ggml-tune.h

@@ -78,6 +78,8 @@ int ggml_mulmat_tune_validate(struct ggml_mulmat_tune *tune,
 int ggml_mulmat_tune_setup_model(struct ggml_mulmat_tune *tune,
                                  const char *model, int m_num);
 
+int ggml_mulmat_tune_free(struct ggml_mulmat_tune *tune);
+
 int ggml_mulmat_tune_write_data(const struct ggml_mulmat_tune *tune, FILE *fp);
 
 int ggml_mulmat_tune_read_data(struct ggml_mulmat_tune *tune, FILE *fp);

ggml.c

@@ -34,7 +34,7 @@
 
 typedef volatile LONG atomic_int;
 typedef atomic_int atomic_bool;
-typedef atomic_int atomic_flag;
+typedef LONG atomic_flag;
 
 static void atomic_store(atomic_int* ptr, LONG val) {
     InterlockedExchange(ptr, val);
@@ -49,10 +49,10 @@ static LONG atomic_fetch_sub(atomic_int* ptr, LONG dec) {
     return atomic_fetch_add(ptr, -(dec));
 }
 
-static inline LONG atomic_flag_test_and_set(atomic_flag* ptr) {
+static inline LONG atomic_flag_test_and_set(volatile atomic_flag* ptr) {
     return InterlockedCompareExchange(ptr, 1, 0);
 }
-static inline LONG atomic_flag_clear(atomic_flag* ptr) {
+static inline LONG atomic_flag_clear(volatile atomic_flag* ptr) {
     return InterlockedExchange(ptr, 0);
 }
 
@@ -14361,7 +14361,7 @@ void ggml_graph_compute_mul_mat_set_task_profile(struct ggml_cgraph *cgraph) {
 
     const int mm_cache_len = 16;
     struct mm_cache_element mm_cache[mm_cache_len];
-    memset(mm_cache, 0, sizeof(mm_cache));
+    memset(mm_cache, 0, sizeof(struct mm_cache_element) * mm_cache_len);
 
     // TODO: optimize if we are sure that the M is a fixed value.
 
@@ -14489,8 +14489,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
     if (n_threads > 1) {
         state_shared = (struct ggml_compute_state_shared){
             .spin = { 0 },
-            .mutex = PTHREAD_MUTEX_INITIALIZER,
-            .cond = PTHREAD_COND_INITIALIZER,
             .n_tasks = 0,
             .n_waiting = 0,
             .wait_now = false,
@@ -14500,13 +14498,18 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
             .stop = false,
         };
 
+        int rc;
+        rc = pthread_mutex_init(&state_shared.mutex, NULL);
+        GGML_ASSERT(rc == 0);
+        rc = pthread_cond_init(&state_shared.cond, NULL);
+        GGML_ASSERT(rc == 0);
+
         size_t sz_workers = sizeof(struct ggml_compute_state) * (n_threads - 1);
         workers = alloca(sz_workers);
         GGML_ASSERT(workers);
         memset(workers, 0, sz_workers);
 
         // NOTE: we could delay creating workers.
-        int rc;
         for (int j = 0; j < n_threads - 1; j++) {
             workers[j].shared = &state_shared;
             rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);

tests/test-mulmat-tune

@@ -165,6 +165,8 @@ void test_ggml_mulmat_tune_estimate_time_non_zero_NK(void) {
         }
     }
 
+    ggml_mulmat_tune_free(&tune);
+
     printf("%2d of %2d pass\n", n_pass, n_tests * shape->n_profiles);
 }
 
@@ -314,5 +316,7 @@ void test_ggml_mulmat_tune_estimate_time_zero_NK(void) {
         }
     }
 
+    ggml_mulmat_tune_free(&tune);
+
     printf("%2d of %2d pass\n", n_pass, n_tests * shape->n_profiles);
 }