GPU abstraction for TBDR optimizations

[K0bin]

Render targets that exist only in tile memory

• Vulkan: VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT + VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT
• Metal: MTLStorageMode.memoryless
• D3D12: Not possible, needs allocation
• WebGPU: Not possible, needs allocation

Render passes

• Vulkan: Dynamic rendering
• Metal: MTLRenderCommandEncoder
• D3D12: Render passes
• WebGPU: Render passes

Multi threading

• Vulkan: Secondary command buffers or VK_RENDERING_SUSPENDING_BIT + VK_RENDERING_RESUMING_BIT
• Metal: MTLParallelRenderCommandEncoder
• D3D12: D3D12_RENDER_PASS_FLAG_SUSPENDING_PASS + D3D12_RENDER_PASS_FLAG_RESUMING_PASS
• WebGPU: Not possible, WebGPU is strictly single threaded

Accessing tile data

• Vulkan: VK_KHR_dynamic_rendering_local_read, VK_EXT_shader_tile_image or (for Qualcomm) VK_QCOM_tile_shading
• Metal: Tile shaders
• D3D12: D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_PRESERVE_LOCAL_SRV + D3D12_RENDER_PASS_BEGINNING_ACCESS_TYPE_PRESERVE_LOCAL_UAV