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

Handle devices with only a single queue

Browse source
This commit is contained in:
0cc4m 2024-01-23 17:34:45 +01:00
parent 1c953c10a0
commit 566a178c8f
1 changed files with 16 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

25
ggml-vulkan.cpp
View file

@ -128,6 +128,7 @@ struct vk_device {
vk_queue compute_queue;
vk_queue transfer_queue;
uint32_t descriptor_set_mode;
bool is_igpu;
};
struct vk_op_push_constants {
@ -1003,6 +1004,7 @@ std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
std::cerr << "ggml_vulkan: Using " << vk_device.properties.deviceName << std::endl;
vk_device.vendor_id = vk_device.properties.vendorID;
vk_device.is_igpu = vk_device.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
bool fp16_storage = false;
bool fp16_compute = false;
@ -1026,15 +1028,17 @@ std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1);
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 float compute_queue_priority = 1.0f;
const float transfer_queue_priority = 1.0f;
const float priorities[] = { 1.0f, 1.0f };
const bool single_queue = compute_queue_family_index != transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
std::vector<vk::DeviceQueueCreateInfo> device_queue_create_infos;
if (compute_queue_family_index != transfer_queue_family_index) {
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, &compute_queue_priority});
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, &transfer_queue_priority});
} else {
const float priorities[] = { compute_queue_priority, transfer_queue_priority };
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1});
} else if(!single_queue) {
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities});
} else {
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
}
vk::DeviceCreateInfo device_create_info;
std::vector<const char *> device_extensions;
@ -1091,10 +1095,13 @@ std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
ggml_vk_load_shaders();
// Queues
const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
vk_device.compute_queue = ggml_vk_create_queue(compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer });
vk_device.transfer_queue = ggml_vk_create_queue(transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer });
if (!single_queue) {
const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
vk_device.transfer_queue = ggml_vk_create_queue(transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer });
} else {
vk_device.transfer_queue = vk_device.compute_queue;
}
vk_fence = vk_device.device.createFence({});