GPU: RTX 5080
OS: Windows 11 x64
Driver version: 595.92
When using VK_EXT_descriptor_heap + VK_DESCRIPTOR_MAPPING_SOURCE_PUSH_ADDRESS_EXT to “bind” uniform buffers, data loaded from those uniform buffers in shaders are loaded using global memory load instructions instead of constant load instructions.
This causes a small perf regression compared to other methods such as vkCmdPushDescriptorSet.
Here is a comparison of what is reported in Nsight Graphics for a shader line where I load a matrix from an uniform buffer with either method:
Push descriptor
Descriptor heap + push address
