happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

use multiple events to avoid reset issues

This commit is contained in:
Ruben Ortlam 2026-03-14 06:39:20 +01:00
parent eebf21c3e9
commit 4374b5ab9a
1 changed files with 41 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -944,12 +944,13 @@ struct vk_semaphore {
uint64_t value; uint64_t value;
}; };
// vk_event is used for the event-related backend interfaces. It uses 'event' for // vk_event is used for the event-related backend interfaces. It uses vk::Events for
// event_wait and a timeline semaphore for event_synchronize. Polling on an event for // event_wait and a timeline semaphore for event_synchronize. Polling on an event for
// event_synchronize wouldn't be sufficient to wait for command buffers to complete, // event_synchronize wouldn't be sufficient to wait for command buffers to complete,
// and would lead to validation errors. // and would lead to validation errors.
struct vk_event { struct vk_event {
vk::Event event; std::vector<vk::Event> events;
size_t next_event_idx;
vk_semaphore tl_semaphore; vk_semaphore tl_semaphore;
vk_command_buffer* cmd_buffer = nullptr; vk_command_buffer* cmd_buffer = nullptr;
uint64_t cmd_buffer_use_counter = 0; uint64_t cmd_buffer_use_counter = 0;
@ -14871,10 +14872,14 @@ static void ggml_backend_vk_event_record(ggml_backend_t backend, ggml_backend_ev
vk_context compute_ctx = ggml_vk_get_compute_ctx(ctx); vk_context compute_ctx = ggml_vk_get_compute_ctx(ctx);
auto* cmd_buf = compute_ctx->s->buffer; // retrieve pointer before it gets reset auto* cmd_buf = compute_ctx->s->buffer; // retrieve pointer before it gets reset
// the backend interface doesn't have an explicit reset, so reset it here // Grab the next event and record it, create one if necessary
// before we record the command to set it if (vkev->next_event_idx == vkev->events.size()) {
ggml_vk_reset_event(compute_ctx, vkev->event); vkev->events.push_back(ctx->device->device.createEvent({}));
ggml_vk_set_event(compute_ctx, vkev->event); }
vk::Event& cur_event = vkev->events[vkev->next_event_idx];
vkev->next_event_idx++;
ggml_vk_set_event(compute_ctx, cur_event);
vkev->tl_semaphore.value++; vkev->tl_semaphore.value++;
compute_ctx->s->signal_semaphores.push_back(vkev->tl_semaphore); compute_ctx->s->signal_semaphores.push_back(vkev->tl_semaphore);
@ -14894,7 +14899,11 @@ static void ggml_backend_vk_event_wait(ggml_backend_t backend, ggml_backend_even
vk_context compute_ctx = ggml_vk_get_compute_ctx(ctx); vk_context compute_ctx = ggml_vk_get_compute_ctx(ctx);
ggml_vk_wait_events(compute_ctx, {vkev->event}); if (vkev->next_event_idx > 0) {
// Wait for latest event
vk::Event& cur_event = vkev->events[vkev->next_event_idx - 1];
ggml_vk_wait_events(compute_ctx, { cur_event });
}
} }
// TODO: enable async and synchronize // TODO: enable async and synchronize
@ -15684,9 +15693,8 @@ static ggml_backend_event_t ggml_backend_vk_device_event_new(ggml_backend_dev_t
return nullptr; return nullptr;
} }
// The event/fence is expected to initially be in the signaled state. // No events initially, they get created on demand
vkev->event = device->device.createEvent({}); vkev->next_event_idx = 0;
device->device.setEvent(vkev->event);
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 }; vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
vk::SemaphoreCreateInfo ci{}; vk::SemaphoreCreateInfo ci{};
@ -15706,7 +15714,9 @@ static void ggml_backend_vk_device_event_free(ggml_backend_dev_t dev, ggml_backe
vk_event *vkev = (vk_event *)event->context; vk_event *vkev = (vk_event *)event->context;
device->device.destroySemaphore(vkev->tl_semaphore.s); device->device.destroySemaphore(vkev->tl_semaphore.s);
device->device.destroyEvent(vkev->event); for (auto& event : vkev->events) {
device->device.destroyEvent(event);
}
delete vkev; delete vkev;
delete event; delete event;
} }
@ -15717,19 +15727,28 @@ static void ggml_backend_vk_device_event_synchronize(ggml_backend_dev_t dev, ggm
auto device = ggml_vk_get_device(ctx->device); auto device = ggml_vk_get_device(ctx->device);
vk_event *vkev = (vk_event *)event->context; vk_event *vkev = (vk_event *)event->context;
vk::Semaphore sem = vkev->tl_semaphore.s; // Only do something if the event has actually been used
uint64_t val = vkev->tl_semaphore.value; if (vkev->next_event_idx > 0) {
vk::SemaphoreWaitInfo swi{vk::SemaphoreWaitFlags{}, sem, val}; vk::Semaphore sem = vkev->tl_semaphore.s;
VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize"); uint64_t val = vkev->tl_semaphore.value;
vk::SemaphoreWaitInfo swi{vk::SemaphoreWaitFlags{}, sem, val};
VK_CHECK(device->device.waitSemaphores(swi, UINT64_MAX), "event_synchronize");
// Finished using current command buffer so we flag for reuse // Reset all events and flag for for reuse
if (vkev->cmd_buffer) { for (size_t i = 0; i < vkev->next_event_idx; i++) {
// Only flag for reuse if it hasn't been reused already device->device.resetEvent(vkev->events[i]);
if (vkev->cmd_buffer_use_counter == vkev->cmd_buffer->use_counter) { }
vkev->cmd_buffer->in_use = false; vkev->next_event_idx = 0;
vkev->cmd_buffer->buf.reset();
// Finished using current command buffer so we flag for reuse
if (vkev->cmd_buffer) {
// Only flag for reuse if it hasn't been reused already
if (vkev->cmd_buffer_use_counter == vkev->cmd_buffer->use_counter) {
vkev->cmd_buffer->in_use = false;
vkev->cmd_buffer->buf.reset();
}
vkev->cmd_buffer = nullptr;
} }
vkev->cmd_buffer = nullptr;
} }
} }