ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 11:21:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reduce backend synchronization calls

Browse Source
This commit is contained in:
Kawrakow
2025-12-09 06:19:50 +00:00
parent 5fe3979951
commit d0fec69966
1 changed files with 25 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
ggml/src/ggml-backend.cpp
View File

@@ -12,6 +12,7 @@
#include <string> #include <string>
#include <vector> #include <vector>
#include <set> #include <set>
#include <array>
#define IK_PRINT_TIMING 0 #define IK_PRINT_TIMING 0
@@ -1879,6 +1880,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
std::vector<uint32_t> unique_ids; std::vector<uint32_t> unique_ids;
ggml_tensor * last_ids_tensor = nullptr; ggml_tensor * last_ids_tensor = nullptr;
std::array<bool, GGML_SCHED_MAX_BACKENDS> needs_sync{{true}};
for (int i = 0; i < sched->n_splits; i++) { for (int i = 0; i < sched->n_splits; i++) {
#if IK_PRINT_TIMING #if IK_PRINT_TIMING
int64_t tim1 = ggml_time_us(); int64_t tim1 = ggml_time_us();
@@ -1897,18 +1900,24 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
if (input->flags & GGML_TENSOR_FLAG_INPUT) { if (input->flags & GGML_TENSOR_FLAG_INPUT) {
// inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
if (sched->events[split_backend_id][sched->cur_copy] != NULL) { if (needs_sync[split_backend_id]) {
ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]); if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
} else { ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
ggml_backend_synchronize(split_backend); } else {
ggml_backend_synchronize(split_backend);
}
needs_sync[split_backend_id] = false;
} }
ggml_backend_tensor_copy(input, input_cpy); ggml_backend_tensor_copy(input, input_cpy);
} else { } else {
// wait for the split backend to finish using the input before overwriting it // wait for the split backend to finish using the input before overwriting it
if (sched->events[split_backend_id][sched->cur_copy] != NULL) { if (needs_sync[split_backend_id]) {
ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]); if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
} else { ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
ggml_backend_synchronize(split_backend); } else {
ggml_backend_synchronize(split_backend);
}
needs_sync[split_backend_id] = false;
} }
ggml_tensor * node = split->graph.nodes[0]; ggml_tensor * node = split->graph.nodes[0];
@@ -2001,16 +2010,20 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
// TODO: add public function to facilitate this, since applications do not have direct access to the backend interface // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) { if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
ggml_backend_synchronize(input_backend); ggml_backend_synchronize(input_backend);
if (sched->events[split_backend_id][sched->cur_copy] != NULL) { if (needs_sync[split_backend_id]) {
ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]); if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
} else { ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
ggml_backend_synchronize(split_backend); } else {
ggml_backend_synchronize(split_backend);
}
needs_sync[split_backend_id] = false;
} }
ggml_backend_tensor_copy(input, input_cpy); ggml_backend_tensor_copy(input, input_cpy);
} }
} }
} }
needs_sync[split_backend_id] = true;
if (!sched->callback_eval) { if (!sched->callback_eval) {
#if IK_PRINT_TIMING #if IK_PRINT_TIMING
int64_t tim2 = ggml_time_us(); int64_t tim2 = ggml_time_us();