Vulkan Memory Report layer: Vulkan Memory by Usage Type - #28
Conversation
Implements VK_LAYER_GOOGLE_DeviceMemoryReport layer providing memory tracking by Usage Type. Exposes VK_EXT_device_memory_report and classifies allocations by Vulkan usage types (surface, texture, buffer, other). Emits Perfetto counters starting with vulkan.mem.app.usage.* and vulkan.mem.driver.usage.*.
olehkuznetsov
left a comment
olehkuznetsov
left a comment
There was a problem hiding this comment.
In Vulkan, applications often allocate a single large block of VkDeviceMemory and bind different sections (ranges) of it to multiple virtual resources (like buffers and images) using offsets. This is done to stay within the driver's maximum allocation count limit.
The current implementation of the memory report layer does not track partial memory bindings (sub-allocations). Instead, it attributes the entire size of the physical allocation to the category of the most recently bound resource.
Done. |
a8e31c1 to
86e0cfa
Compare
Also ensure the Vulkan layers would be built if only the memory report was enabled.
Populate physical device to instance mapping during vkCreateInstance to ensure logical device creation succeeds even if the application creates a device without explicitly calling vkEnumeratePhysicalDevices.
Query extension properties via instance_dispatch_table(physicalDevice) directly rather than relying on instance singleton map lookup.
Move std::call_once inside InitializeDeviceMemoryReportPerfetto to ensure thread-safe Perfetto initialization and remove the static once_flag from the header.
…ocation Pre-define static CounterTrack instances for known Vulkan memory report categories to eliminate heap allocation overhead in GetCounterTrack during tracing hot paths.
Remove the handwritten implementation and export of vkEnumerateDeviceExtensionProperties from the DeviceMemoryReport layer, allowing extension enumeration to be handled natively via the loader dispatch table and layer manifest.
Introduce helper functions to classify Vulkan buffer and image usage flags and physical memory property flags into refined memory usage clusters (staging_transfer, ray_tracing, indirect_gpu_driven, compute_storage, uniform_constants, geometry_mesh, transient_memoryless, etc.) with unit tests covering classification precedence.
…esource tracking Capture and cache VkPhysicalDeviceMemoryProperties on vkCreateDevice, resolve memory property flags via memoryTypeIndex in vkAllocateMemory, and refactor Resource metadata to resolve usage clusters dynamically during memory binding and unbound headroom updates.
emrekultursay
left a comment
emrekultursay
left a comment
There was a problem hiding this comment.
LGTM. Please wait for Oleh's approval to merge.
4 participants
Implements VK_LAYER_GOOGLE_DeviceMemoryReport layer providing memory tracking by Usage Type. Exposes VK_EXT_device_memory_report and classifies allocations by Vulkan usage types (surface, texture, buffer, other). Emits Perfetto counters starting with vulkan.mem.app.usage.* and vulkan.mem.driver.usage.*.