nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 19 Packages Projects Releases Wiki Activity

vulkan: use graphics queue on AMD (#20551)

Browse Source 
* vulkan: use graphics queue on AMD for slightly better performance

* disable async transfer queue on AMD
This commit is contained in:
Ruben Ortlam
2026-03-15 08:18:54 +01:00
committed by GitHub
parent 6b10a82c00
commit 1a3d8edbba
1 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+5 -5
View File
@@ -4981,8 +4981,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties(); std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();
// Try to find a non-graphics compute queue and transfer-focused queues // Try to find a non-graphics compute queue and transfer-focused queues
const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1); // On AMD, the graphics queue seems to be faster, so don't avoid it
const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1); const vk::QueueFlagBits graphics_flag = device->vendor_id == VK_VENDOR_ID_AMD ? (vk::QueueFlagBits)0 : vk::QueueFlagBits::eGraphics;
const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, graphics_flag, -1, 1);
const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | graphics_flag, compute_queue_family_index, 1);
const float priorities[] = { 1.0f, 1.0f }; const float priorities[] = { 1.0f, 1.0f };
device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1; device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
@@ -5441,13 +5443,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
ggml_vk_load_shaders(device); ggml_vk_load_shaders(device);
const bool prefers_transfer_queue = device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != AMD_GCN;
if (!device->single_queue) { if (!device->single_queue) {
const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0; const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true); ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
device->async_use_transfer_queue = prefers_transfer_queue || (getenv("GGML_VK_ASYNC_USE_TRANSFER_QUEUE") != nullptr); device->async_use_transfer_queue = (getenv("GGML_VK_ASYNC_USE_TRANSFER_QUEUE") != nullptr);
} else { } else {
// TODO: Use pointer or reference to avoid copy // TODO: Use pointer or reference to avoid copy
device->transfer_queue.copyFrom(device->compute_queue); device->transfer_queue.copyFrom(device->compute_queue);