GPU

Thin Vulkan 1.3 abstraction. Does not hide Vulkan — it codifies the patterns mjolnir relies on (bindless texture array, slab-allocated mesh buffers, dynamic rendering, frame-in-flight semaphores). Most user code never touches this directly; it exists for render and as an escape hatch.

Design choices

• No render-graph, no auto-barrier inference. Passes record their own barriers. Easier to reason about, easier to debug in RenderDoc, no surprise hazards.
• Dynamic rendering only. No VkRenderPass / VkFramebuffer objects. Attachments come from begin_rendering per pass.
• Bindless first. Every sampled texture goes through one big descriptor array (TextureManager); every mesh goes through one vertex / index / skinning buffer (MeshManager). Draw calls pass indices, not descriptor sets.
• Pre-baked pipeline state. Common rasterizer / blend / depth states are exposed as ALL_CAPS constants so pipelines compose declaratively instead of repeating boilerplate.

Key subsystems

• GPUContext — instance, device, queues, descriptor pool, command pools. One per engine.
• Swapchain — frames-in-flight fences and semaphores, presentation.
• TextureManager — bindless arena for 2D + cube textures. Allocations return a Texture2DHandle / TextureCubeHandle (a generational index into the descriptor array).
• MeshManager — three immutable GPU buffers (vertex / index / skinning) sub-allocated via slab allocators. Mesh allocation returns stable byte offsets that draw-indirect commands reference directly.
• MutableBuffer(T) / ImmutableBuffer(T) — typed wrappers around host-visible vs device-local memory. MutableBuffer is mapped for CPU writes; ImmutableBuffer.write uploads via a one-shot staging buffer.

Frames in flight

FRAMES_IN_FLIGHT = 2 by default. Per-frame resources (semaphores, per-camera attachments, staging buffers) are indexed by frame_index % FRAMES_IN_FLIGHT. Anything the GPU might still be reading must outlive that window — see architecture for the staging-age rule.