vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

silence robot

Browse source
This commit is contained in:
zrm 2023-05-23 17:08:37 -04:00
parent 9d058c2096
commit 2c1b5ae197
2 changed files with 17 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

26
ggml.c
View file

@ -505,7 +505,7 @@ struct ggml_numa_nodes ggml_numa = {
void ggml_numa_init(void) void ggml_numa_init(void)
{ {
if (ggml_numa.n_nodes > 0) return; if (ggml_numa.n_nodes > 0) { return; }
#ifdef __linux__ #ifdef __linux__
struct stat st; struct stat st;
char path[256]; char path[256];
@ -514,17 +514,21 @@ void ggml_numa_init(void)
while (true) { while (true) {
rv = snprintf(path, sizeof(path), "/sys/devices/system/node/node%u", ggml_numa.n_nodes); rv = snprintf(path, sizeof(path), "/sys/devices/system/node/node%u", ggml_numa.n_nodes);
GGML_ASSERT(rv > 0 && (unsigned)rv < sizeof(path)); GGML_ASSERT(rv > 0 && (unsigned)rv < sizeof(path));
if (stat(path, &st) != 0) break; if (stat(path, &st) != 0) { break; }
++ggml_numa.n_nodes; ++ggml_numa.n_nodes;
} }
// enumerate CPUs // enumerate CPUs
while (true) { while (true) {
rv = snprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%u", ggml_numa.total_cpus); rv = snprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%u", ggml_numa.total_cpus);
GGML_ASSERT(rv > 0 && (unsigned)rv < sizeof(path)); GGML_ASSERT(rv > 0 && (unsigned)rv < sizeof(path));
if (stat(path, &st) != 0) break; if (stat(path, &st) != 0) { break; }
++ggml_numa.total_cpus; ++ggml_numa.total_cpus;
} }
GGML_PRINT_DEBUG("found %u numa nodes, %u CPUs\n", ggml_numa.n_nodes, ggml_numa.total_cpus); GGML_PRINT_DEBUG("found %u numa nodes, %u CPUs\n", ggml_numa.n_nodes, ggml_numa.total_cpus);
if (ggml_numa.n_nodes < 1 || ggml_numa.total_cpus < 1) {
ggml_numa.n_nodes = 0;
return;
}
ggml_numa.nodes = calloc(ggml_numa.n_nodes, sizeof(struct ggml_numa_node)); ggml_numa.nodes = calloc(ggml_numa.n_nodes, sizeof(struct ggml_numa_node));
GGML_ASSERT(ggml_numa.nodes != NULL); GGML_ASSERT(ggml_numa.nodes != NULL);
for (uint32_t n = 0; n < ggml_numa.n_nodes; ++n) { for (uint32_t n = 0; n < ggml_numa.n_nodes; ++n) {
@ -14058,18 +14062,18 @@ typedef pthread_t ggml_thread_t;
#ifdef __linux__ #ifdef __linux__
void set_numa_thread_affinity(int thread_n, int n_threads) void set_numa_thread_affinity(int thread_n, int n_threads)
{ {
if (!ggml_is_numa()) return; if (!ggml_is_numa()) { return; }
// run thread on node_num thread_n / (threads per node) // run thread on node_num thread_n / (threads per node)
int node_num = thread_n / (n_threads / ggml_numa.n_nodes); int node_num = thread_n / (n_threads / ggml_numa.n_nodes);
struct ggml_numa_node *node = &ggml_numa.nodes[node_num]; struct ggml_numa_node *node = &ggml_numa.nodes[node_num];
size_t setsize = CPU_ALLOC_SIZE(ggml_numa.total_cpus); size_t setsize = CPU_ALLOC_SIZE(ggml_numa.total_cpus);
cpu_set_t *cpus = CPU_ALLOC(ggml_numa.total_cpus); cpu_set_t *cpus = CPU_ALLOC(ggml_numa.total_cpus);
CPU_ZERO_S(setsize, cpus); CPU_ZERO_S(setsize, cpus);
for (size_t i=0; i < node->n_cpus; ++i) { for (size_t i = 0; i < node->n_cpus; ++i) {
CPU_SET_S(node->cpus[i], setsize, cpus); CPU_SET_S(node->cpus[i], setsize, cpus);
} }
int rv; int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
if ((rv = pthread_setaffinity_np(pthread_self(), setsize, cpus))) { if (rv) {
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",
strerror(rv)); strerror(rv));
} }
@ -14077,15 +14081,15 @@ void set_numa_thread_affinity(int thread_n, int n_threads)
} }
void clear_numa_thread_affinity(void) void clear_numa_thread_affinity(void)
{ {
if (!ggml_is_numa()) return; if (!ggml_is_numa()) { return; }
size_t setsize = CPU_ALLOC_SIZE(ggml_numa.total_cpus); size_t setsize = CPU_ALLOC_SIZE(ggml_numa.total_cpus);
cpu_set_t *cpus = CPU_ALLOC(ggml_numa.total_cpus); cpu_set_t *cpus = CPU_ALLOC(ggml_numa.total_cpus);
CPU_ZERO_S(setsize, cpus); CPU_ZERO_S(setsize, cpus);
for (unsigned i=0; i < ggml_numa.total_cpus; ++i) { for (unsigned i = 0; i < ggml_numa.total_cpus; ++i) {
CPU_SET_S(i, setsize, cpus); CPU_SET_S(i, setsize, cpus);
} }
int rv; int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
if((rv = pthread_setaffinity_np(pthread_self(), setsize, cpus))) { if (rv) {
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",
strerror(rv)); strerror(rv));
} }

4
llama-util.h
View file

@ -180,9 +180,9 @@ struct llama_mmap {
int fd = fileno(file->fp); int fd = fileno(file->fp);
int flags = MAP_SHARED; int flags = MAP_SHARED;
// prefetch/readahead impairs performance on NUMA systems // prefetch/readahead impairs performance on NUMA systems
if (ggml_is_numa()) prefetch = 0; if (ggml_is_numa()) { prefetch = 0; }
#ifdef __linux__ #ifdef __linux__
if (prefetch) flags |= MAP_POPULATE; if (prefetch) { flags |= MAP_POPULATE; }
#endif #endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
if (addr == MAP_FAILED) { if (addr == MAP_FAILED) {