CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: initialize vulkan-hpp to allow using extension function pointers (#15705)

Browse Source 
Use this to query register count for shader compiles on NVIDIA. Currently
this is only for performance debug, but it could eventually be used in some
heuristics like split_k.
This commit is contained in:
Jeff Bolz
2025-09-13 16:23:30 +01:00
committed by GitHub
parent 50f4281a6f
commit b9c9c9f789
1 changed files with 43 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -5,8 +5,14 @@
#include "ggml-cpu.h"
#endif
// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers-
#define VULKAN_HPP_DISPATCH_LOADER_DYNAMIC 1
#include <vulkan/vulkan.hpp>
// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers-
VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE
#include <algorithm>
#include <cmath>
#include <iomanip>
@@ -121,6 +127,8 @@ struct vk_pipeline_struct {
bool needed {};
// set to true when the shader has been compiled
bool compiled {};
// number of registers used, extracted from pipeline executable properties
uint32_t register_count {};
};
typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline;
@@ -429,6 +437,8 @@ struct vk_device_struct {
bool coopmat2;
bool pipeline_executable_properties_support {};
size_t idx;
bool mul_mat_l[GGML_TYPE_COUNT];
@@ -1603,6 +1613,20 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast<VkDebugUtilsObjectNameInfoEXT &>(duoni));
}
if (device->pipeline_executable_properties_support) {
vk::PipelineExecutableInfoKHR executableInfo;
executableInfo.pipeline = pipeline->pipeline;
auto statistics = device->device.getPipelineExecutableStatisticsKHR(executableInfo);
for (auto & s : statistics) {
// "Register Count" is reported by NVIDIA drivers.
if (strcmp(s.name, "Register Count") == 0) {
VK_LOG_DEBUG(pipeline->name << " " << s.name << ": " << s.value.u64 << " registers");
pipeline->register_count = (uint32_t)s.value.u64;
}
}
}
{
std::lock_guard<std::recursive_mutex> guard(device->mutex);
device->all_pipelines.push_back(pipeline);
@@ -3610,6 +3634,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
bool amd_shader_core_properties2 = false;
bool pipeline_robustness = false;
bool coopmat2_support = false;
bool pipeline_executable_properties_support = false;
device->coopmat_support = false;
device->integer_dot_product = false;
bool bfloat16_support = false;
@@ -3652,6 +3677,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
!getenv("GGML_VK_DISABLE_BFLOAT16")) {
bfloat16_support = true;
#endif
} else if (strcmp("VK_KHR_pipeline_executable_properties", properties.extensionName) == 0) {
pipeline_executable_properties_support = true;
}
}
@@ -3878,8 +3905,18 @@ static vk_device ggml_vk_get_device(size_t idx) {
device_extensions.push_back("VK_KHR_shader_integer_dot_product");
}
VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR pep_features {};
pep_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
if (pipeline_executable_properties_support) {
last_struct->pNext = (VkBaseOutStructure *)&pep_features;
last_struct = (VkBaseOutStructure *)&pep_features;
device_extensions.push_back("VK_KHR_pipeline_executable_properties");
}
vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
device->pipeline_executable_properties_support = pipeline_executable_properties_support;
device->fp16 = device->fp16 && vk12_features.shaderFloat16;
#if defined(VK_KHR_shader_bfloat16)
@@ -4395,6 +4432,9 @@ static void ggml_vk_instance_init() {
}
VK_LOG_DEBUG("ggml_vk_instance_init()");
// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers-
VULKAN_HPP_DEFAULT_DISPATCHER.init(vkGetInstanceProcAddr);
uint32_t api_version = vk::enumerateInstanceVersion();
if (api_version < VK_API_VERSION_1_2) {
@@ -4462,6 +4502,9 @@ static void ggml_vk_instance_init() {
vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
// See https://github.com/KhronosGroup/Vulkan-Hpp?tab=readme-ov-file#extensions--per-device-function-pointers-
VULKAN_HPP_DEFAULT_DISPATCHER.init(vk_instance.instance);
std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
// Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan