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

fix: ggml: make GGML compatible with vulkan v1.2.162

Browse source 
Add the option GGML_VULKAN_V1_2_162 to make GGML
compatible with vulkan v1.2.162.
This option is OFF by default.

vulkan-header: v1.2.162
https://github.com/KhronosGroup/Vulkan-Headers/tree/v1.2.162
This commit is contained in:
Junil Kim 2024-11-22 12:39:04 +09:00
parent 755a9b2bf0
commit 6c4720c4a7
2 changed files with 29 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

5
CMakeLists.txt
View file

@ -73,6 +73,11 @@ option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
# 3rd party libs
option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
option(GGML_VULKAN_V1_2_162 "llama: make GGML compatible with vulkan v1.2.162" OFF)
if (GGML_VULKAN_V1_2_162 AND GGML_VULKAN)
add_definitions(-DGGML_VULKAN_V1_2_162)
endif()
# Required for relocatable CMake package
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)

24
ggml/src/ggml-vulkan.cpp
View file

@ -65,6 +65,13 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA
#define VK_LOG_DEBUG(msg) ((void) 0)
#endif // GGML_VULKAN_DEBUG
#ifdef GGML_VULKAN_V1_2_162
#ifdef VK_NULL_HANDLE
#undef VK_NULL_HANDLE
#define VK_NULL_HANDLE nullptr
#endif // VK_NULL_HANDLE
#endif // GGML_VULKAN_V1_2_162
struct ggml_backend_vk_context;
struct vk_queue {
@ -865,6 +872,16 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
tl_signal_vals[idx].push_back(submission.signal_semaphores[i].value);
tl_signal_semaphores[idx].push_back(submission.signal_semaphores[i].s);
}
#if defined(GGML_VULKAN_V1_2_162)
vk::TimelineSemaphoreSubmitInfo timeline_info(
(uint32_t) submission.wait_semaphores.size(),
tl_wait_vals[idx].data(),
(uint32_t) submission.signal_semaphores.size(),
tl_signal_vals[idx].data()
);
timeline_info.setPNext(nullptr);
tl_submit_infos.push_back(timeline_info);
#else
tl_submit_infos.push_back({
(uint32_t) submission.wait_semaphores.size(),
tl_wait_vals[idx].data(),
@ -873,6 +890,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
});
tl_submit_infos[idx].sType = vk::StructureType::eTimelineSemaphoreSubmitInfo;
tl_submit_infos[idx].pNext = nullptr;
#endif
vk::SubmitInfo si{
(uint32_t) submission.wait_semaphores.size(),
tl_wait_semaphores[idx].data(),
@ -1846,13 +1864,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
vk::PhysicalDeviceProperties2 props2;
vk::PhysicalDeviceMaintenance3Properties props3;
#ifndef GGML_VULKAN_V1_2_162
vk::PhysicalDeviceMaintenance4Properties props4;
#endif // GGML_VULKAN_V1_2_162
vk::PhysicalDeviceSubgroupProperties subgroup_props;
props2.pNext = &props3;
props3.pNext = &subgroup_props;
#ifndef GGML_VULKAN_V1_2_162
if (maintenance4_support) {
subgroup_props.pNext = &props4;
}
#endif // GGML_VULKAN_V1_2_162
device->physical_device.getProperties2(&props2);
device->properties = props2.properties;
@ -1860,8 +1882,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
#ifndef GGML_VULKAN_V1_2_162
} else if (maintenance4_support) {
device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize);
#endif // GGML_VULKAN_V1_2_162
} else {
device->max_memory_allocation_size = props3.maxMemoryAllocationSize;
}