From: Winston Ma <redacted>
Date: Mon, 1 Jun 2026 12:03:32 +0000 (+0800)
Subject: vulkan: reduce host memory lock contention (#23376)
X-Git-Tag: upstream/0.0.10438~981
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=bef69f1306285b95ff65fbd726f9dc4958f205d0;p=pkg%2Fggml%2Fsources%2Fllama.cpp

vulkan: reduce host memory lock contention (#23376)

* vulkan: reduces lock contention

* replace unique_lock with lock_guard
---

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 3cf191f20..c3d4c7a71 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -62,6 +62,7 @@ typedef struct VkPhysicalDeviceCooperativeMatrixDecodeVectorFeaturesNV {
 #include <map>
 #include <set>
 #include <unordered_map>
+#include <shared_mutex>
 #include <mutex>
 #include <future>
 #include <thread>
@@ -618,6 +619,7 @@ static constexpr std::initializer_list<std::array<int, 3>> rms_norm_mul_rope_vie
 
 struct vk_device_struct {
     std::recursive_mutex mutex;
+    mutable std::shared_mutex pinned_memory_mutex;
 
     vk::PhysicalDevice physical_device;
     vk::PhysicalDeviceProperties properties;
@@ -7010,7 +7012,7 @@ static void * ggml_vk_host_malloc(vk_device& device, size_t size) {
         return nullptr;
     }
 
-    std::lock_guard<std::recursive_mutex> guard(device->mutex);
+    std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
     device->pinned_memory.push_back(std::make_tuple(buf->ptr, size, buf));
 
     return buf->ptr;
@@ -7021,7 +7023,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
         return;
     }
     VK_LOG_MEMORY("ggml_vk_host_free(" << ptr << ")");
-    std::lock_guard<std::recursive_mutex> guard(device->mutex);
+    std::lock_guard<std::shared_mutex> guard(device->pinned_memory_mutex);
 
     vk_buffer buf;
     size_t index;
@@ -7045,7 +7047,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
 }
 
 static void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
-    std::lock_guard<std::recursive_mutex> guard(device->mutex);
+    std::shared_lock<std::shared_mutex> guard(device->pinned_memory_mutex);
     buf = nullptr;
     buf_offset = 0;
     for (size_t i = 0; i < device->pinned_memory.size(); i++) {