wmantly/llama.cpp-xrip
1
0 Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

vulkan: fix amd iGPU slow checkpoint read (#30049)

Browse Source
This commit is contained in:
Ruben Ortlam authored and GitHub committed 2026-10-07 08:20:04 +02:00
1 parent 78651c410d
commit 4d756bc72b
1 file changed
+4 -1
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan-buffers.cpp
+4 -1
View File
@@ -623,7 +623,10 @@ void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, size_t sp
// If the device is not an UMA device the memory is host-accessible through rebar. While writing
// through PCIe is sufficient fast reading back data from PCIe is slower than going through
// the HW device to host copy path.
if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
// AMD UMA: uncached host-visible memory is write-combined, CPU reads are slow
const bool slow_host_read = src->device->vendor_id == VK_VENDOR_ID_AMD &&
!(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCached);
if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma && !slow_host_read) {
GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
std::lock_guard<std::recursive_mutex> guard(src->device->mutex);